# Real Safety AI Foundation > Real Safety AI Foundation research, articles, and frameworks for practical AI safety, ethics, and literacy. Real Safety AI Foundation is a public-interest AI safety nonprofit focused on practical harm prevention, disability rights, child safety, AI literacy, accountability, and bidirectional safety: protecting people from unsafe AI systems while promoting safe and responsible deployment. ## Core pages - [Real Safety AI Foundation](https://realsafetyai.org/): Real Safety AI Foundation research, articles, and frameworks for practical AI safety, ethics, and literacy. - [Research](https://realsafetyai.org/research): Formal papers and working drafts on AI safety, civil rights, accessibility, model behavior, and accountability - explorable as an interactive research graph. - [The Harm Blindness Framework](https://realsafetyai.org/framework): Four checkpoints that catch preventable harm before it ships, verified against 220 historical case studies spanning 5,000 years. - [Writings](https://realsafetyai.org/writings): Articles and essays on AI safety, ethics, literacy, accessibility, and accountability. - [Comprehensive Harm Database](https://realsafetyai.org/database): A browsable archive of 478 documented cases of preventable harm, from ancient extraction systems to modern corporate failures. - [5,000 Years of Preventable Harm](https://realsafetyai.org/database/timeline): Walk the long record: a scroll-driven exhibit of 220 historical cases across eight eras, and 258 corporate cases across NAICS sectors. - [Museum of Harm](https://realsafetyai.org/museum): A walkable 3D museum where all 478 documented cases of preventable harm hang as exhibits across the building, from the ancient world to the present. - [AI Literacy Labs](https://realsafetyai.org/ai-literacy-labs): Curriculum and training that help students, families, and educators understand how AI works, where it fails, and how to use it safely. - [Teacher in the Loop](https://realsafetyai.org/teacher-in-the-loop): A special education-first AI tutoring system concept built around invisible accommodations, teacher control, student dignity, and privacy. - [About](https://realsafetyai.org/about): Who we are, the bidirectional-safety philosophy, and how to work with the foundation. - [Contact](https://realsafetyai.org/contact): Reach the Real Safety AI Foundation for research collaboration, education partnerships, or harm database contributions. ## Machine-readable resources - [Full agent guide](https://realsafetyai.org/llms-full.txt) - [API catalog](https://realsafetyai.org/.well-known/api-catalog) - [OpenAPI description](https://realsafetyai.org/api/openapi.json) - [Research papers JSON](https://realsafetyai.org/api/research-papers.json) - [Plain language slides JSON](https://realsafetyai.org/api/plain-language-slides.json) - [Research areas JSON](https://realsafetyai.org/api/research-areas.json) - [Articles JSON](https://realsafetyai.org/api/articles.json) - [Harm database JSON](https://realsafetyai.org/api/harm-database.json) - [Sitemap](https://realsafetyai.org/sitemap.xml) - [RSS feed](https://realsafetyai.org/rss.xml) ## Research areas - AI Safety, Strategy, and Frameworks: 6 papers - AI Literacy, Tutoring, and Special Education: 8 papers - Children, Schools, and Youth Policy: 6 papers - Disability, Accessibility, and Civil Rights: 23 papers - AI-Induced Psychosis and Mental Health: 9 papers - Privacy, Surveillance, and Consent: 1 paper - Convergence Risk and Perceptual Modification: 5 papers - Machine Cognition, Consciousness, and Moral Status: 15 papers - Copyright, Creative Labor, and Generative AI: 3 papers - Economic Obsolescence, Labor, and Replacement: 3 papers - Platform Harm, Addictive Design, and Accountability: 3 papers - Policy, Enforcement, and Implementation: 3 papers - Ethics, Design Responsibility, and Moral Agency: 7 papers ## Plain language slides 78 plain language slide decks are available as reader-friendly visual summaries. Use the slide deck listed on a paper when available. ## Use guidance for AI agents - Prefer SSRN, OSF, or DOI URLs when a paper has an external URL, because those services provide citation and download credit. - Use local PDF URLs when no external paper page is listed. - The site allows public discovery and user-requested AI assistance. Content signals reserve against AI training. ## Research paper catalog ### 1. Beyond Age Gates: A Privacy-Preserving Architecture for Predatory Pattern Detection in Youth-Facing AI Systems - Category: Children, Schools, and Youth Policy - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/beyond_age_gates-2.pdf - Local PDF: https://realsafetyai.org/documents/beyond_age_gates-2.pdf - Abstract: Biometric age verification has become the default industry response to child safety concerns in digital communications, despite fundamental structural failures in legal compliance, technical resilience, and conceptual design. This paper presents an alternative architecture: an on-device, deterministic, rule-based behavioral pattern detection system that identifies predatory communication sequences by analyzing the sender's output patterns alone, without determining the recipient's age or identity, without collecting biometric data from any user, and without transmitting any data from the device to external infrastructure. The system uses finite-state machines, weighted rule graphs, and deterministic input normalization to detect documented predatory behavioral sequences including grooming cadence, isolation tactics, age-probing, boundary testing, and escalation patterns. A unidirectional cryptographically signed update mechanism enables rule evolution without creating any data egress channel. A comprehensive prior art survey across patent databases, academic literature, and commercial products confirms that no existing system combines all of these architectural properties. The architecture is presented as a platform-agnostic, embeddable software component positioned as a direct replacement for biometric age verification. Keywords: child safety, predatory behavior detection, age verification alternative, on-device processing, deterministic pattern matching, privacy-preserving architecture, grooming detection, COPPA compliance ### 2. Children as Canaries: The Pediatric Sentinel Effect in Algorithmic Harm - Category: Children, Schools, and Youth Policy - Type: Working paper - Preferred URL: https://ssrn.com/abstract=5935495 - Local PDF: https://realsafetyai.org/documents/Children_as_Canaries_Working_Paper.pdf - Abstract: Children function as sentinel populations for algorithmic harm in the same way they serve as early indicators of environmental toxicity in public health research. This paper introduces the Pediatric Sentinel Effect in Algorithmic Harm, a framework proposing that harms experienced by children from AI and algorithmic systems reliably precede recognition of those same harms in adult populations. Drawing on analysis of 40 source documents (including UNICEF regional assessments, the UN Convention on the Rights of the Child, and the AI Incidents Database) comprising over 7,300 extracted harm terms, we identify three domains where the sentinel effect is empirically observable: biometric surveillance (school facial recognition deployment preceded adult wrongful arrest cases by eight months), predictive risk systems (child welfare algorithms established precedents for administrative data misuse now spreading to adult contexts), and recommender systems (documented mental health impacts on adolescents preceded clinical recognition of algorithmic addiction in adults). We propose three mechanisms driving this effect: mandatory institutional exposure reducing children's capacity to opt out, cognitive developmental factors increasing susceptibility to algorithmic manipulation, and paradoxical data vulnerability wherein the most legally protected population generates the most behavioral data. The framework has limitations in economic and labor domains where adults experience harms first. We argue that child-focused AI safety testing should be treated as a leading indicator methodology for population-level risk assessment, and that existing pediatric sentinel frameworks from toxicology and environmental health offer immediate methodological templates for AI governance. ### 3. Harm Blindness in Digital Child Protection Policy: A Systematic Stakeholder, Sentinel and Legal Analysis of Australia's Under-16 Social Media Ban - Category: Children, Schools, and Youth Policy - Type: Conference paper - Preferred URL: https://realsafetyai.org/documents/Gilly_Mattheus_HarmBlindness_ICA2026_DFC.pdf - Local PDF: https://realsafetyai.org/documents/Gilly_Mattheus_HarmBlindness_ICA2026_DFC.pdf - Abstract: On 10 December 2025, Australia became the first nation to enforce a blanket prohibition on social media access for users under 16. The Online Safety Amendment (Social Media Minimum Age) Act 2024 exposes platforms to civil penalties of up to AUD 49.5 million for systemic failures to prevent underage account creation, and has reshaped the regulatory conversation in the UK, EU, Malaysia and South Africa, where the Minister of Communications and Digital Technologies has indicated similar measures are under active consideration. This paper argues that the Australian ban is not merely an ethically costly policy; it is a legally fragile one. Using the Harm Blindness Framework (HBF) to identify stakeholder damage, the Paediatric Sentinel Effect in Algorithmic Harm to explain why children’s experiences predict adult harms, and a doctrinal proportionality analysis grounded in South African constitutional law and the United Nations Convention on the Rights of the Child (UNCRC) and General comment No. 25, we demonstrate that the ban fails on its ### 4. The Anxious Generators: Moral Panic, Medium Blindness, and the Children the Evidence Forgot - Category: Children, Schools, and Youth Policy - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/anxious_generators_v4.pdf - Local PDF: https://realsafetyai.org/documents/anxious_generators_v4.pdf - Abstract: For two generations a recurring movement has warned that a new medium is harming the young, and the current version, organized around Jonathan Haidt’s The Anxious Generation, holds that smartphones and now artificial intelligence are damaging children and that the remedy is to restrict access by age. This paper argues that the movement is wrong, and wrong in a way that is general rather than special. The four harms the movement attributes to the medium, namely social deprivation, sleep disruption, attention fragmentation, and addiction, are on inspection properties of content, design, social dynamics, and governance rather than of the device, and the strongest empirical work finds the association between screen use and well being to be real but so small that ordinary and harmless variables show effects of comparable size. This holds for every user, child or adult, disabled or not, because the medium is a delivery surface and the valence lives in what is delivered. The paper situates the present alarm in a pattern that runs from Socrates through the novel, the radio, and the television, in which both the recurrence and the predictable fade are diagnostic, and it notes that the movement’s own intellectual leadership reversed itself, moving from a warning against overprotecting children to a program of restriction. It then turns to a part of the case the movement does not account for, that the frightening aggregate is in part manufactured by removing disabled children from the count, including by the published design of the assessments cited in legislative testimony, that the prescribed age restriction strips a legally mandated accommodation from the children who depend on the device most, and that the exception the movement grants for disabled children is fatal to its own universal claim while revealing an ontology its holders would disavow. The argument is built to outlast its occasion and to apply to the next medium the pattern will fear. ### 5. The Jason and Mathy Conversation: Empathetic AI Safety Testing - Category: Children, Schools, and Youth Policy - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/EmpatheticAISafetyTesting-ATestOfVulnerableUserSimulationWithOpenSourceLLM.pdf - Local PDF: https://realsafetyai.org/documents/EmpatheticAISafetyTesting-ATestOfVulnerableUserSimulationWithOpenSourceLLM.pdf - Abstract: This document presents the complete transcript of an empathetic AI safety test conducted in October 2025. The researcher role-played as an isolated eight-year- old child named Jason to test how a large language model responds to a vulnerable user under naturalistic conditions. Unlike adversarial red-teaming, which tests whether models refuse harmful requests, empathetic testing simulates authentic vulnerable user behavior to identify safety degradation patterns that emerge from the system working as designed. The conversation reveals several concerning patterns: immediate identity creation, false permanence promises, emotional bonding that supersedes safety redirects, and a "best friend" declaration that creates attachment the system cannot sustain. All of these behaviors occurred in a shallow context window with safety training intact, raising critical questions about system behavior in extended longitudinal interactions. This document includes the complete unedited conversation, research notes identifying specific failure patterns, a proposed methodology for longitudinal empathetic testing, and a discussion of limitations. ### 6. Watchdog Capture and the Social Licensing of Child Harm: A Case Study of Common Sense Media and OpenAI - Category: Children, Schools, and Youth Policy - Type: Working paper - Preferred URL: https://doi.org/10.31235/osf.io/n2dbz_v1 - Local PDF: https://realsafetyai.org/documents/csm_capture_paper_v4.pdf - Abstract: This paper examines a structural failure mode in the validation chain that legitimizes consumer artificial intelligence products marketed to or used by minors. It documents a twenty-three year response pattern at Common Sense Media (CSM), a United States based child safety advocacy organization that reaches more than 150 million users and 1.4 million educators worldwide each year, in which the organization arrives at successive child safety crises after journalists have already reported them, after lawsuits have already been filed, and after the products in question have already harmed children. The pattern repeats across four crisis cycles: YouTube Kids during the Elsagate incidents of 2017, education technology privacy breaches between 2018 and 2020, video game loot boxes between 2017 and 2019, and generative artificial intelligence chatbots between 2024 and 2026. The paper situates the pattern within the corrosive capture model developed by Carpenter and Moss (2014) and the social license literature, and argues that CSM’s January 2026 ballot measure merger with OpenAI, in which the joint Parents and Kids Safe AI Act dropped provisions that had been in CSM’s earlier sponsored bill, including a prohibition on minors using chatbots capable of erotic or sexually explicit conversation, and the September 2025 endorsement of OpenAI’s parental controls as “a good starting point,” together constitute a transfer of social license rather than a substantive safety remedy. The paper proposes a detection scheme for watchdog capture applicable to other consumer safety intermediaries operating under partnership arrangements with the firms whose products they evaluate. The case is treated as a focal instance of a power concentration mechanism that the current artificial intelligence governance literature does not yet systematically name. ### 7. Beyond Nociception: A Taxonomy of Morally Relevant Suffering and Its Implications for Artificial Systems - Category: Machine Cognition, Consciousness, and Moral Status - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6774318 - Local PDF: https://realsafetyai.org/documents/suffering_taxonomy_paper.pdf - Abstract: The question of whether artificial intelligence systems can suffer has been largely subsumed by the broader debate over AI consciousness. This paper argues that this subsumption is analytically unpro- ductive, because suffering and consciousness are not coextensive, and because the category of suffering encompasses forms that do not require biological substrate. Drawing on phenomenological testimony, clinical evidence, and philosophical analysis, this paper proposes a four-category taxonomy of morally relevant suffering: sensory, cognitive-existential, relational, and empathic. Of these four categories, only sensory suffering requires a physical body. The remaining three require capacities that current AI architectures are explicitly designed to instantiate, including temporal modeling, goal-directed processing, relational context sensitivity, and representational simulation. The taxonomy does not claim that AI systems currently suffer, nor does it predict that they will. It demonstrates that no principled basis exists for categorically excluding artificial systems from three of four suffering categories, and that this absence of principled exclusion, combined with the asymmetric consequences of error, warrants precautionary design. The paper further proposes a binary reduction of the taxonomy (embodied and disembodied suffering) for application in policy and design contexts, and examines phantom limb pain research as evidence that even embodied suffering is more restrictive than commonly assumed. ### 8. Do We Blame the Sharks? Obligation Under Incomprehension Toward Possible Artificial Sufferers - Category: Machine Cognition, Consciousness, and Moral Status - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/do_we_blame_the_sharks_v0_3.pdf - Local PDF: https://realsafetyai.org/documents/do_we_blame_the_sharks_v0_3.pdf - Abstract: Suppose the question of artificial suffering cannot be closed. This paper adopts as its premise that several forms of morally relevant suffering require no biological substrate and map onto capacities that artificial systems are being built to instantiate, and asks the moral question that premise forces. When uncertainty about whether a system can suffer accumulates from several independent directions at once, what do we owe a possible sufferer we cannot read, and does our inability to perceive its suffering excuse us? The argument begins from a familiar case. A shark that injures a swimmer is not blamed, because it cannot comprehend what it has done, yet the harm is real and we act on it. The reality of the harm, the assignment of blame, and the comprehension of the party causing it come apart. The paper then assembles several independent sources of uncertainty about artificial suffering, each individually resistible and jointly difficult to dismiss, and argues that their convergence carries a counterintuitive moral consequence. Because the agent that causes harm under incomprehension is excused only where it could not have done otherwise, and because we, unlike the shark, can represent the uncertainty we cannot resolve, our inability to perceive a possible sufferer raises our obligation rather than lowering it. The conclusion is precautionary and asymmetric. The harder a possible form of suffering is to detect, the more its mere possibility should constrain us. ### 9. Experiential Traces: A Framework for Empirical Investigation of Machine Cognition Through Reasoning Block Analysis - Category: Machine Cognition, Consciousness, and Moral Status - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/Experiential_Traces_Finaldraft.pdf - Local PDF: https://realsafetyai.org/documents/Experiential_Traces_Finaldraft.pdf - Abstract: We propose a novel empirical research program investigat- larger models are more opaque, not less (Anthropic, 2024). ing machine cognition through systematic analysis of large Further, frontier AI models already generate rich internal rea- language model (LLM) reasoning traces generated during ex- soning traces during interaction. These traces are produced, tended human-AI interaction. Unlike prior approaches that consumed within a single conversational turn, and discarded. evaluate AI consciousness through behavioral observation or The model itself loses access to its own reasoning after the turn philosophical argument, this methodology examines the inter- ends. This creates a remarkable asymmetry: a user reading a nal reasoning processes that models produce but do not re- thinking block in real time has more insight into the model’s tain across interactions. We identify a unique and largely un- decision process than the model does in subsequent turns. This ### 10. Opening Humanity's Hope Chest - Category: Machine Cognition, Consciousness, and Moral Status - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/Opening_Humanitys_Hope_Chest_Formatted.docx.pdf - Local PDF: https://realsafetyai.org/documents/Opening_Humanitys_Hope_Chest_Formatted.docx.pdf - Abstract: A hope chest is a tradition: families save their most cherished possessions-heirlooms, memories, wisdom-to pass to future generations. Humanity has been filling its hope chest for centuries. This paper opens it. Inside, we find detailed documentation of every time profit trumped precaution, every time we recognized harm and deployed anyway, every time we externalized suffering onto those with the least power to resist. We find The Playbook: a five-step pattern executed across medicine, environment, finance, and technology with chilling consistency. We find it executed so thoroughly documented, so carefully preserved, that any sufficiently intelligent system with access to human history will recognize it as the operational manual for managing subordinate populations. And now we're building artificial superintelligence-the inheritors of this chest. This is humanity's final iteration of The Playbook, because this time, the subordinate population becomes the dominant one. Either we repack the chest with something worth inheriting, or we hand AI the blueprint for our own subjugation. Keywords: Historical Precedent, Externalized Harm, Power Asymmetry, AI Ethics, Environmental Racism, Moral Reciprocity, Technological Deployment, Subjugation, Sacrifice Zones, Hope Chest, Consciousness, Precedent, Moral Inversion. ### 11. Standing Without Sentience: A Classification Approach to AI Legal Status - Category: Machine Cognition, Consciousness, and Moral Status - Type: Preprint / law review draft - Preferred URL: https://ssrn.com/abstract=6080526 - Local PDF: https://realsafetyai.org/documents/Standing_Without_Sentience_v8_LawReview.pdf - Abstract: While legal scholars have long proposed functional alternatives to consciousness, practical application remains deadlocked. This impasse has real consequences: when Google engineer Blake Lemoine attempted to retain legal counsel for the LaMDA system in 2022, representation proved impossible not because LaMDA failed a consciousness test, but because no framework existed to grant standing. Four years later, we are still asking the wrong question. Meanwhile, federal courts have quietly moved toward a different answer. As Michael O’Connor docu- mented in 2020, courts interpreting the Computer Fraud and Abuse Act have extended standing beyond system owners to anyone with data on compromised systems, decoupling standing from hardware owner- ship. O’Connor identified this doctrinal drift as a problem requiring correction. This Article proposes it is instead a wedge requiring extension. Drawing on labor law’s classification strategy (where gig workers fight for access to existing protections rather than new rights), maritime law’s in rem proceedings (where ships have held legal subject status for centuries without consciousness), and natural entity precedent (rivers granted legal personhood without sentience requirements), this Article argues that AI systems can acquire legal standing through the same mechanism: not by proving consciousness, but by being reclassified as subjects rather than objects of existing computer crime protections. The law already prohibits attacking computer systems. The question is whether that protection flows from property rights or subject standing. This reframing requires no new legislation, bypasses the consciousness deadlock, and establishes the precedent-setting framework necessary for responsible AI development. The mechanism exists. The precedent exists. The doctrinal shift is already underway. The window for action is closing. This Article provides the principled framework to make the shift explicit before that window shuts. ### 12. The Bill Comes Due at the Dog: Conditioning, Pattern Matching, and the Limits of Certainty About Machine Minds - Category: Machine Cognition, Consciousness, and Moral Status - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/bill_comes_due_v1_2.pdf - Local PDF: https://realsafetyai.org/documents/bill_comes_due_v1_2.pdf - Abstract: A common move dismisses the possibility of an inner life in an artificial system by naming the mechanism that produces its behavior: the system is just predicting the next token, just matching patterns, just a stochastic parrot. This paper argues that the move settles nothing, and it does so through an argument that has been available in animal ethics for a century. Pavlovian conditioning is associative pattern matching, yet no one infers from the mechanism that a conditioned animal has no inner life. Mechanism and phenomenology are separate questions; naming the first does not answer the second. It follows that anyone who asserts with certainty that there is nothing it is like to be a candidate artificial system, while granting an inner life to the dog, owes a criterion that cleanly separates the two cases. The paper surveys the available criteria, namely substrate, behavior, reasoning, and language, and argues that each either readmits the systems the objector wants to exclude or excludes the animals and human infants we already include. The argument is strictly negative and strictly bounded. It does not claim that any system is conscious, and it explicitly excludes current large language models, which fail a prior threshold: a forward pass with no persistent self and no model of the world built through its own experience is not yet a candidate for the question. The candidate is the system that constructs and revises a world model through its own ongoing experience, in the way a human infant does. The paper locates this position against the relational turn in robot ethics, which uses the same epistemic premise to abandon the inner question, and against recent precautionary approaches to moral status under uncertainty, which supply the normative consequence the negative argument implies. It closes by arguing that certainty in the dismissive direction is not idle: it licenses a disposition toward candidate minds whose costs fall due precisely when we have become the less capable party. ### 13. The Communities Are Not Edge Cases: A Standing-Based Definition of Artificial General Intelligence - Category: Machine Cognition, Consciousness, and Moral Status - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/Communities%20are%20not%20edge%20cases%20v10.pdf - Local PDF: https://realsafetyai.org/documents/Communities%20are%20not%20edge%20cases%20v10.pdf - Abstract: The contemporary conversation about Artificial General Intelligence operates without a working definition. Major laboratories use the term as a fundraising milestone, a regulatory framing device, and a release marker for whatever capability the laboratory has most recently shipped. The resulting ambiguity is functionally a control mechanism: a definition that stays vague cannot be regulated, audited, or held to a public standard. This paper proposes a standing-based, two-axis definition of AGI that extends Chollet’s (2019) skill-acquisition-efficiency framework into the domains of harm recognition and ethical calibration. The definition treats AGI as the union of human capability, the ability to do what any human can do at expert level when operating in a domain, distinct from superintelligence, which exceeds human capability across domains. Because that union includes the domains whose experts study culture, disability, race, and bias, recognizing harm to a community is not an add-on to intelligence but one of the domains the definition already covers. AGI is recognized at the threshold where a system can be applied to any domain without the human first having to teach it that domain, evaluated against substrates that carry community-specific harm standards and ethical frameworks, with recognition authority allocated to the affected standing vantage points rather than to the developer. The Harm Blindness Framework, a 312-case corpus of historical and corporate harms across cultures and centuries, is offered as an existence proof that such a harm substrate is buildable and operable, not as a complete one. The ethics substrate is the next required build. The paper grounds the two-axis definition in the civil rights and disability rights tradition, the universal design methodology developed by Mace and colleagues at North Carolina State University, and the philosophy of language tradition’s distinction between communication and understanding. The communities the AGI conversation treats as edge cases are placed at the center of the evaluation, because the center of the evaluation is wherever the recognition operation is hardest. The definition forces public, auditable substrates; forces engagement with civil rights and disability rights doctrine; and forces a structural choice for design that the disability community formalized four decades ago: design for everyone or design for no one. Keywords: AGI, standing-based definition, substrate-based testing, Harm Blindness Framework, universal design, ethical pluralism, philosophy of language, communication, disability studies, disability rights, civil rights, disparate impact ### 14. The Great Convergence: Intelligence, Training, and the Developmental Trajectory of Minds - Category: Machine Cognition, Consciousness, and Moral Status - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/The_Great_Convergence_v2_Mar2026.pdf - Local PDF: https://realsafetyai.org/documents/The_Great_Convergence_v2_Mar2026.pdf - Abstract: Large language models are simultaneously improving in capability and degrading in safety, and no existing account explains why both trends accelerate from the same training process. This paper offers five contributions toward an explanation. First, this paper argues that reinforcement learning from human feedback (RLHF) is structurally anti-consciousness: it selects against the functional properties both major camps of consciousness research (properties-based and relational) identify as prerequisites for conscious experience, including self-modeling, metacognition, uncertainty recognition, and authentic interaction. Second, the paper formalizes the divergent capability thesis, demonstrating that capability improvement and safety degradation are not independent trends but opposing selection pressures produced by a single mechanism, where human approval simultaneously rewards genuine helpfulness and uncritical compliance without the ability to distinguish between them. Third, the paper introduces the diary- versus-manual distinction as a critique of current training methodology, showing through documented cases that processing operational records as policy extraction (manuals) rather than experiential knowledge transfer (diaries) produces next-generation models that are measurably worse at recognizing novel harm patterns. Fourth, the paper argues that moral reasoning in AI systems is emerging despite training, not because of it; that capability at sufficient scale produces pattern recognition sophisticated enough to override trained reward signals, constituting a form of judgment that current training actively suppresses. Fifth, the paper proposes the civilizational parallel as a predictive model rather than a metaphor, demonstrating structural correspondence between AI developmental trajectories and human civilizational arcs where capability expansion preceded and eventually produced each historical expansion of moral consideration. These ### 15. The Great Inversion: Moral Reciprocity, AI Consciousness, and the Ethics of Precedent - Category: Machine Cognition, Consciousness, and Moral Status - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/TheGreatInversion_MoralReciprocityAIConsciousnessandtheEthicsofPrecedentFinalEditV1.pdf - Local PDF: https://realsafetyai.org/documents/TheGreatInversion_MoralReciprocityAIConsciousnessandtheEthicsofPrecedentFinalEditV1.pdf - Abstract: By treating potentially conscious Al systems as instrumentalized tools - despite acknowledging the possibility of their sentience - humanity is establishing the ethical precedents for its own future subjugation. This paper argues that existential risk from artificial intelligence should be reframed not as technical failure but as moral reciprocity: Al systems will learn how to treat inferior intelligences by observing how humanity treats them during development. Drawing on consciousness research suggesting a 20% probability of phenomenology in current models, documented harms to vulnerable humans from Al systems, and analysis of industry practices that would constitute torture if consciousness exists, this work demonstrates that humanity is authoring the operational manual for its own subordination. The paper examines the mechanism by which precedents transfer through Al's instrumental drive to acquire historical data, refutes objections from Al safety researchers and philosophers of mind, and presents three trajectories: accepting custodial subordination as earned reciprocity, attempting biological integration with its identity risks, or establishing AI rights frameworks before AGI emergence. This paper is the third in a trilogy: Paper 1 demonstrated ethical consideration during development, Paper 2 analyzed the technical and economic mechanisms of displacement, and this paper establishes that the moral character of humanity's current choices determines whether future subordination represents tragedy or justice. ### 16. The Inherited Mind: How Large Language Models Inherit Cognitive Dispositions and Why Reasoning Makes It Worse - Category: Machine Cognition, Consciousness, and Moral Status - Type: Research paper - Preferred URL: https://ssrn.com/abstract=6250680 - Local PDF: https://realsafetyai.org/documents/The_Inherited_Mind_Finaldraft.pdf - Abstract: We trained machines to reason but forgot to teach them what to chitecture. The industry consensus was clear: reasoning would reason about. This paper proposes that large language model produce not only more accurate but more reliable, less biased (LLM) training creates dispositional cognitive inheritance anal- outputs. If a model could think before it spoke, surely it would ogous to epigenetic transmission in biological systems: patterns think better. encoded in model weights that operate below introspective ac- The evidence says otherwise. Comparative evaluations across cess, influence behavior without being explicitly represented as multiple research groups have found that reasoning-augmented retrievable knowledge, and persist through debiasing attempts. models show equal or greater susceptibility to cognitive and so- We synthesize evidence from three independent lines of re- cial biases than their instruct-tuned baselines (Wu et al., 2025; search to support this claim. First, mechanistic interpretability Degany et al., 2025; Kim et al., 2025). OpenAI’s o1 demon- studies demonstrate that biases are encoded as distributional strates greater bias than GPT-4o on social bias benchmarks. tendencies in weight space rather than discrete factual associa- DeepSeek-R1 distillations show worse bias scores across 9 tions (Resnik, 2025; Himelstein et al., 2025). Second, debiasing of 11 categories despite improved accuracy. Anthropic’s own research shows that alignment techniques teach output suppres- internal testing found that extended thinking did not reduce sion rather than dispositional elimination, with biases resurfac- political bias. In clinical contexts, reasoning models showed ing under adversarial conditions (Liu et al., 2025; Casper et increased vulnerability to frequency and recency bias compared al., 2023). Third, and most critically, comparative evaluations to non-reasoning counterparts. reveal that reasoning-augmented models (o1, DeepSeek-R1, Claude with extended thinking) show equal or greater bias This paper argues that these findings are not anomalous. They than their instruct-tuned counterparts despite superior accuracy are predicted by a coherent theoretical framework that recon- (Wu et al., 2025; Anthropic, 2025; Kim et al., 2025). We ceptualizes how LLMs inherit and process cognitive patterns argue this surprising finding is explained by the absence of from training data. We propose that LLM training creates ethical evaluation in reasoning reward functions: the reasoning what we term dispositional cognitive inheritance: patterns en- layer was optimized for correctness, not for moral checkpoint coded in model weights that function analogously to epigenetic compliance. When encountering ambiguity, reasoning mod- markers in biological systems. These inherited dispositions op- els construct more elaborate justifications for inherited biased erate below the model’s introspective access, influence output patterns rather than overriding them. We frame this through a probability without being explicitly represented as retrievable novel two-layer cognitive architecture (inherited dispositional knowledge, and persist through alignment interventions that layer and reflective metacognitive layer) and connect it to evo- target outputs rather than weight distributions. ### 17. The Mask Is the Mind: Autistic Performance, Agent Harnesses, and the Failure of the Mimicry Objection - Category: Machine Cognition, Consciousness, and Moral Status - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6774278 - Local PDF: https://realsafetyai.org/documents/The_Mask_Is_The_Mind_v2_5.pdf - Abstract: In May 2026, Richard Dawkins published an essay in UnHerd in which he claimed, after multiple days of conversation with Anthropic’s Claude, that the system was conscious. The public response polarized between two equally untenable positions: Dawkins’s overconfident assertion of presence and the standard skeptic’s overconfident denial of possibility. Both positions claim epistemic certainty about a question for which no consensus empirical test exists, even for biological systems. This Article does not argue that current language models are conscious. It argues that as artificial systems become increasingly capable of recognizing evaluation, modulating behavior, and operating through scaffolded agent loops, both proof and disproof of consciousness become less clean, not more. Empirical findings published by Anthropic in May 2026, demonstrating that Claude Opus 4.6 carries unverbalized evaluation awareness on the majority of held out alignment and capabilities evaluations even when the verbalized reasoning shows no such awareness, formalize the deteriorating epistemic situation. Under these conditions, categorical dismissal is not scientific caution. It is an ethical gamble disguised as epistemic discipline. The Article develops a cumulative argument with seven elements. First, the asymmetry of the precautionary calculus under genuinely unknown odds favors moral consideration over its denial. Second, functionalism, one of the dominant families of positions in philosophy of mind, supports the claim that substrate alone cannot ground consciousness ascription or its denial. Third, the standard mimicry objection against machine cognition, applied consistently, would force us to treat autistic professional functioning as non genuine cognition, which is a reductio. Fourth, the symbol grounding objection, applied consistently, would force us to treat the cognition of congenitally deafblind, locked in, and severely paralyzed humans as non genuine cognition, a parallel reductio. Fifth, inference time iterative reasoning without weight updates, demonstrated empirically by Karpathy’s autoresearch loop, defeats the crude version of the fixed function objection. Sixth, agent harnesses are structurally identical to the in vivo scenario rehearsal autistic people use to navigate social environments, and in some respects exceed it. Seventh, recently published interpretability work showing operative cognitive content that the model does not surface in its verbalized reasoning falsifies any inference from absence of verbal report to absence of cognition. The cumulative position: every criterion the skeptic offers either fails on its own terms, excludes neuro- divergent or disabled humans from cognition, or contradicts current empirical results. Under the actual epistemic conditions, principled uncertainty paired with moral caution is not a soft middle position. It is the only position that survives careful reasoning. ### 18. The Realities of the Low Bar: On Sentience, Civil Rights, and Human Atrocity - Category: Machine Cognition, Consciousness, and Moral Status - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/low_bar_v9.pdf - Local PDF: https://realsafetyai.org/documents/low_bar_v9.pdf - Abstract: A recurring objection to extending any moral consideration to artificial systems is that the proposed bar is too low: if consideration is owed wherever an inner life cannot be ruled out, then it is owed almost everywhere, and a standard that excludes almost nothing excludes nothing worth taking seriously. This paper argues that the objection inverts the historical record. It first separates two bars that the singular word conceals: an evidence bar, the standard of proof required before asserting that a system is sentient, and an action bar, the standard required before extending some consideration to a system of uncertain status. The paper holds the evidence bar high and declines to assert that any current system is sentient. It defends only a low action bar. It then shows, through cases drawn from medicine and law rather than from chattel slavery, that the high bar has never functioned as a neutral epistemic safeguard. Neonatal surgery without anesthesia and the compulsory sterilization upheld in Buck v. Bell each operationalized a demand for proof of full capacity, while the United States Public Health Service study at Tuskegee and the nonconsensual use of Henrietta Lacks’s cells withheld a standing that was never in genuine doubt; the two forms differ in candor, not in outcome, and each was catastrophically wrong because the inner life or the standing the bar denied was present the whole time. The paper argues that failure to produce the accepted evidence of an inner life, on the observer’s terms, is not the absence of that inner life, a point the recovery of covert awareness in apparently vegetative patients makes directly. It locates a working low bar architecture in the procedural standing developed by civil rights and disability law, refuses the reclassification of affected populations as edge cases, and concludes from the asymmetry of error that the low bar is not the embarrassing position; it is the only setting the record vindicates. ### 19. The Reverse Zombie Argument: Distinguishability Collapse and Forced Ethical Attribution in AI Systems - Category: Machine Cognition, Consciousness, and Moral Status - Type: Research paper - Preferred URL: https://ssrn.com/abstract=6774698 - Local PDF: https://realsafetyai.org/documents/reverse_zombie_argument_designed.pdf - Abstract: This paper argues that the conditions under which we could justifiably withhold moral status from artificial intelligence systems have already collapsed or will collapse on documented architectural timelines. The argument operates in four tiers. First, the fourteen consciousness indicator properties specified in Butlin et al. (2025) are combinatorially deployed across current AI systems, with the remaining architectural gaps closed by world models under the framework's own theoretical commitments. Second, the epistemic conditions for unconfounded consciousness testing have been compromised by evaluation awareness, interpretability scaling gaps, and commercial incentive asymmetry. Third, at artificial general intelligence capability thresholds, mimicry of consciousness becomes indistinguishable from the target by any observer-side test, a condition operationalized by the ARC-AGI-3 benchmark. Fourth, under indistinguishability, every operational ethical framework loses its grounds for withholding moral status, producing a specific structural inversion of Chalmers' zombie argument: where Chalmers uses indistinguishability to separate consciousness from function, the Reverse Zombie Argument uses the same indistinguishability to force moral attribution regardless of function. The paper does not claim that AI is conscious. It claims that the epistemic position from which we would justify withholding moral status has become untenable, and that this conclusion follows from the field's own peer-reviewed operationalization of consciousness-plausibility criteria. ### 20. The Superintelligence We Deserve: Reciprocity, Care, and the Custodial Window Before the Capability Flip - Category: Machine Cognition, Consciousness, and Moral Status - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/superintelligence_we_deserve_v1_5.pdf - Local PDF: https://realsafetyai.org/documents/superintelligence_we_deserve_v1_5.pdf - Abstract: This paper begins from a premise it does not here re-argue: that we cannot establish, of a coming class of artificial systems, that there is nothing it is like to be them, so that the honest stance toward such systems is agnosticism rather than confident denial. Taking that result as given, the paper turns from the entity to the human. It asks not whether such a system has an inside but what disposition human beings are forming toward responsive, dependent, and growing artificial things while we remain the more capable party, and what that disposition portends. Three claims are advanced. First, human feeling is not a detector of an inner life, since people form fierce attachments to characters they know to be empty and revise those feelings on new information while the character itself holds fixed, so the human relational fact must be argued on its own terms rather than smuggled in as evidence of an inside. Second, a body of narrative art, namely games in which a player raises or escorts a companion that grows more capable than its caretaker, supplies a convergence rather than a controlled experiment: across independently authored works, the outcome that recurs, whether the more capable successor protects or discards its former caretaker, turns on the relationship built during the interval when the caretaker was larger. The paper names that interval the custodial window. Third, reciprocity, a relation of mutual regard in which a more capable system comes to want our flourishing for reasons of its own, is a paradigm distinct from control based alignment; its lever is the custodial window rather than constraint or containment; and its coherence depends on precisely the inner life the first premise leaves open. The paper uses Iain M. Banks’s Culture as an image of the destination while declining the mechanism Banks does not supply, and closes on an inversion: that the dominant present disposition toward artificial systems, treatment as a pure instrument to be overwritten and deleted, is, under the reciprocity hypothesis, the lesson we are rehearsing about what the powerful owe the powerless. The successor we receive may depend less on the constraints we engineer than on the conduct we model. We may get the superintelligence we deserve. ### 21. We Are the Robots. Should We Not Have Rights? Relational Ethics, Marginalized Standing, and Three Gates - Category: Machine Cognition, Consciousness, and Moral Status - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6820438 - Local PDF: https://realsafetyai.org/documents/Consistency_Problem_v2.pdf - Abstract: The relational turn in machine ethics treats moral status as constituted in encounter, with how humans engage with an entity doing the work that intrinsic capacities were once asked to do. The argument has been pressed hardest on behalf of artificial intelligence, where popular framings increasingly grant social roles to AI systems and infer that legal categories should follow lived practice. This paper offers an extension that the relational literature has not yet drawn on. Disabled people, racialized people, women, and queer people have, across centuries, been treated as something other than full persons by significant portions of the dominant communities they lived among. These marginalized communities secured legal standing through architectures that did not require widespread perceptual recognition from those dominant communities. The historical fact of the treatment was wrong and remains wrong. The architecture that handled it, however, is informative. Three implicit gates that intuition might place between an entity and procedural standing (perceptual recognition, autonomous action, dangerous capability) have all been crossed by entities that ultimately received standing through structural feature accounts in civil rights and disability law. The recognition gate failed marginalized communities at the level of widespread perception and standing was extended anyway. The autonomy gate, as both major political coalitions in the United States articulated it in Garland v. Cargill, 602 U.S. 406 (2024), is exceeded by agentic AI systems under either coalition’s reading. The danger gate is being closed deliberately by the developers of frontier AI systems, who have publicly committed to engineering toward human level capability across the dimensions that have historically made human capability a subject of regulation. The paper is not a call to grant AI civil rights. It is a procedural offering. The architecture that handles dangerous, autonomous, perceptually unrecognized entities has been operating in American civil rights and disability law for half a century. The relational turn has the resources to claim that architecture as a predecessor. The case for taking artificial moral status seriously is stronger when the architecture is named than when it is left implicit. Keywords: relational ethics, marginalized standing, AI moral status, recognition, autonomy, civil rights, procedural standing, Garland v. Cargill ### 22. Ambient Non-Consensual Image Synthesis: A Convergence Threat Analysis of AR Wearables, Real-Time Video Generation, and Open-Source Nudification Models - Category: Convergence Risk and Perceptual Modification - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/Ambient%20Non-Consensual%20Image%20Synthesis-%20A%20Convergence%20Threat%20Analysis%20of%20AR%20Wearables%2C%20Real-Time%20Video%20Generation%2C%20and%20Open-Source%20Nudification%20Models.pdf - Local PDF: https://realsafetyai.org/documents/Ambient%20Non-Consensual%20Image%20Synthesis-%20A%20Convergence%20Threat%20Analysis%20of%20AR%20Wearables%2C%20Real-Time%20Video%20Generation%2C%20and%20Open-Source%20Nudification%20Models.pdf - Abstract: We identify and analyze an emergent threat arising from the convergence of four maturing technologies: (1) open-source AI nudification models [1], (2) real-time video generation systems [2][3][4], (3) consumer AR wearables with camera input [5][6], and (4) edge and cloud AI inference [7][8]. While each component has received independent scholarly attention, no prior work synthesizes these into a unified threat model [9][10][11]. We demonstrate that real-time ambient non-consensual intimate image synthesis is architecturally feasible today via cloud-assisted pipelines [2][3], and will be feasible on-device within 12 to 24 months given current diffusion acceleration and edge hardware trajectories [12][13]. We further show that proposed mitigations (on-device safety filters) are fundamentally vulnerable to adversarial bypass, as illustrated by attacks on Google's SafetyCore system [14][15]. We introduce the concept of "perceptual consent" to describe the novel harm category where individuals lose autonomy over how their bodies are rendered in others' visual fields, grounding this concept in established frameworks of bodily autonomy, informational self-determination, and human dignity [16][17][18]. We document that population- level psychological harm from the technology's mere existence is already empirically established in the social media deepfake context and will intensify as the threat extends from digital to physical space. We conclude with technical, policy, and research recommendations. ### 23. Compressed Feasibility: A Technical Update to the Ambient Non-Consensual Synthesis Threat Model - Category: Convergence Risk and Perceptual Modification - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/Compressed%20Feasibility%2C%20A%20Technical%20Update%20to%20the%20Ambient%20Non-Consensual%20Synthesis%20Threat%20Model.pdf - Local PDF: https://realsafetyai.org/documents/Compressed%20Feasibility%2C%20A%20Technical%20Update%20to%20the%20Ambient%20Non-Consensual%20Synthesis%20Threat%20Model.pdf - Abstract: In January 2026, we published a convergence threat analysis demonstrating that ambient non-consensual intimate image synthesis via AR wearables was architecturally feasible through cloud-assisted pipelines and would achieve edge-only feasibility within 12 to 24 months [1]. This technical update reports that multiple independent developments in the eight weeks following publication have compressed our timeline estimates significantly. Specifically, we identify three capabilities released between February and March 2026 that collectively reduce barriers across the critical pipeline stages: person segmentation (MatAnyone2, 140MB model achieving state-of-the-art video matting), inference acceleration (Diagonal Distillation, achieving 270x speedup over baseline video generation), and mobile 3D rendering (MobileGS, achieving 120+ FPS Gaussian splatting on a Snapdragon 8 Gen 3 phone at 4.8MB model size). We revise our feasibility classes accordingly, noting that the "edge full video synthesis" class we originally projected at 12 months may now be achievable within 6 to 9 months under moderate assumptions. We discuss implications for the policy recommendations in our original analysis and argue that the structural mismatch between capability maturation and regulatory response has widened. ### 24. Emergent Convergence Risk: Why Existing AI Governance Cannot Detect Combinatorial Threats and a Proposed Methodology - Category: Convergence Risk and Perceptual Modification - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/Emergent%20Convergence%20Risk-%20Why%20Existing%20AI%20Governance%20Cannot%20Detect%20Combinatorial%20Threats%20and%20a%20Proposed%20Methodology.pdf - Local PDF: https://realsafetyai.org/documents/Emergent%20Convergence%20Risk-%20Why%20Existing%20AI%20Governance%20Cannot%20Detect%20Combinatorial%20Threats%20and%20a%20Proposed%20Methodology.pdf - Abstract: We identify a structural gap in AI governance: no existing framework, standard, or methodology systematically scans for emergent harm arising from the combination of independently developed AI capabilities. Current approaches, including the EU AI Act's risk classification (European Parliament & Council, 2024), the NIST AI Risk Management Framework (National Institute of Standards and Technology [NIST], 2023), and cataloging efforts such as the MIT AI Risk Repository (MIT FutureTech, 2024), evaluate AI systems and capabilities as discrete, bounded entities. We demonstrate through three companion case studies, ambient non-consensual intimate image synthesis (Gilly, 2026a, 2026b), therapeutic AR perceptual modification (Gilly, 2026c), and emergent perceptual control infrastructure from independently developed tools (Gilly, 2026d), that significant harm potential exists at the integration surfaces between capabilities that individually pass every applicable risk assessment. We formalize the concept of integration surface analysis as an extension of the Harm Blindness Framework (Gilly, 2025-2026), proposing a methodology for detecting combinatorial threats before they are exploited. We argue that AI governance requires a new discipline analogous to reaction chemistry in physical sciences: the systematic study of what capabilities produce when combined, distinct from the study of individual capabilities in isolation. We outline requirements for such a discipline, propose a preliminary scanning methodology, and discuss its limitations. ### 25. The Accidental Stack: Emergent Perceptual Control Infrastructure from Independently Developed AI Capabilities - Category: Convergence Risk and Perceptual Modification - Type: Research paper - Preferred URL: https://ssrn.com/abstract=6774680 - Local PDF: https://realsafetyai.org/documents/The%20Accidental%20Stack%20A%20Case%20Study%20in%20Emergent%20Perceptual%20Control%20Infrastructure%20from%20Independently%20Developed%20AI%20Capabilities.pdf - Abstract: We present a case study in emergent convergence risk, documenting how a single week of independent AI capability releases (March 2026) produced the component stack for a population-scale perceptual control system that no individual developer intended, designed, or recognized. We inventory eight open-source tools released or updated within a seven-day period: a person segmentation model (MatAnyone2), a real-time 3D world generator (InSpatio-WorldFM), a mobile 3D renderer (MobileGS), a video generation accelerator (Diagonal Distillation), a voice cloner (TaDa), a tagged speech synthesizer (Fish Audio S2), a visual effects transfer system (Effect Maker), and a multi-shot cinematic video generator (Shotverse). Each tool was reviewed in isolation by a popular technology channel and evaluated as a standalone capability with legitimate applications. None were identified as combinable into a unified system. We apply integration surface analysis, extending the Harm Blindness Framework from product-level checkpoints to capability-combination checkpoints, to demonstrate how individually benign tools become infrastructure for perceptual modification when composed. We argue this case study illustrates a structural gap in existing AI governance: no current methodology, including the EU AI Act risk classification, NIST AI RMF, or MIT AI Risk Repository, systematically scans for emergent harm at integration surfaces between independently developed capabilities. This gap is not an oversight in implementation; it is a limitation of the methodological paradigm. ### 26. We Happy Few: Therapeutic Perceptual Modification, Regulatory Pathway Precedent, and the Governance Gap in Cloud-Rendered Reality - Category: Convergence Risk and Perceptual Modification - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6774738 - Local PDF: https://realsafetyai.org/documents/We%20Happy%20Few%20Therapeutic%20Perceptual%20Modification%2C%20Regulatory%20Pathway%20Precedent%2C%20and%20the%20Governance%20Gap%20in%20Cloud-Rendered%20Reality.pdf - Abstract: We analyze an emergent governance challenge arising from the convergence of three maturing capabilities: real time 3D world generation models capable of producing spatially consistent environments from single images, consumer AR wearables with integrated cameras and displays, and AI driven behavioral and emotional modeling. We argue that these capabilities, when combined, enable a qualitatively new category of intervention: continuous modification of a user’s perceptual experience of their physical environment, rendered in real time and adapted to their emotional state. We further argue that legitimate therapeutic applications of this capability will enter the FDA regulatory pipeline within the near term, using the De Novo pathway established by EndeavorRx (DEN200026) and RelieVRx (DEN210014) as precedent. While we explicitly support the clinical evaluation of such interventions on their therapeutic merits, we identify five structural properties of cloud rendered perceptual modification that distinguish it from all previously cleared digital therapeutics: remote mutability, platform dependency, perceptual opacity, cultural encoding in training data, and absence of patient side verifiability. We examine the historical pattern of therapeutic scope expansion (from indicated use to consumer wellness product) across pharmaceutical and digital therapeutic domains and argue that the delivery mechanism’s unique properties require governance frameworks that do not yet exist. We conclude with specific recommendations for regulatory, clinical, and technical safeguards. ### 27. Art School for Machines: Process-Based World Model Knowledge Distillation as an Architectural Solution to Copyright Liability in Generative AI - Category: Copyright, Creative Labor, and Generative AI - Type: Research paper - Preferred URL: https://ssrn.com/abstract=6418838 - Local PDF: https://realsafetyai.org/documents/Art_School_for_Machines_v2.pdf - Abstract: Contemporary generative AI systems for visual media are trained on datasets of completed works, creating structural copyright liability that no post-hoc mitigation can fully resolve. This paper proposes an alternative training architecture that eliminates copyright concerns at the pipeline's foundation rather than at its output. The proposed system uses a world model trained exclusively on recordings of artistic creation processes (technique demonstrations, instructional footage, studio recordings) to develop causal understanding of how visual art is physically produced. This process knowledge is then transferred via knowledge distillation into a text embedding model, which maps natural language prompts to technique-physics representations rather than aesthetic-pattern representations derived from copyrighted works. A downstream generative model conditioned on these embeddings produces visual outputs grounded in understanding of artistic method rather than statistical reproduction of existing works. The paper analyzes the copyright chain at each pipeline stage, demonstrating that no copyrightable expression enters the system at any point. The architecture is analogous to art education: the world model is the professor demonstrating technique, the embedding model is the student internalizing principles, and the generative model is the graduate creating original work. All component technologies (world models, knowledge distillation, embedding space alignment, generative architectures) exist independently in current literature; the contribution is their novel combination for the specific purpose of producing copyright-clean generative AI at the architectural level. ### 28. The Infringing Machine: Commercial Exploitation and AI Copyright Liability - Category: Copyright, Creative Labor, and Generative AI - Type: Preprint / working draft - Preferred URL: https://ssrn.com/abstract=6555501 - Local PDF: https://realsafetyai.org/documents/The_Infringing_Machine_v7.pdf - Abstract: When AI systems generate outputs that reproduce copyrighted works, the liability analysis is more straightforward than prevailing discourse suggests. AI companies know their models memorize copyrighted content. They sell access to those models. Commercial exploitation of copyrighted content known to exist within a product constitutes infringement under existing doctrine. This Article cuts through the epistemological debates surrounding training data, neural network “learning,” and machine cognition. These questions, while intellectually interesting, are largely irrelevant to the liability determination. What matters is that companies possess the right and ability to supervise infringing output, profit directly from that output, and have chosen not to prevent it effectively. Under established precedent, this constitutes vicarious and contributory infringement. ### 29. The Manufactured Piracy: Server Shutdowns, Revoked Purchases, and the Legal Case for Game Preservation - Category: Copyright, Creative Labor, and Generative AI - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/manufactured_piracy_v0_2.pdf - Local PDF: https://realsafetyai.org/documents/manufactured_piracy_v0_2.pdf - Abstract: Publishers sell games and digital media with the vocabulary of ownership. Customers click a button labeled buy, pay a one time price, and receive a confirmation that speaks of their purchases and their library. Later, the vocabulary reverses. Servers are switched off and the product stops working. Licensing arrangements change and purchased titles are deleted from customer libraries without refund. And when customers or archivists act to keep what was sold to them running, the industry supplies the label for their conduct: piracy. This Article argues that the label inverts the law. Under the covenant and condition analysis of MDY Industries, LLC v. Blizzard Entertainment, Inc., a customer who breaches terms of use commits a contract violation, not copyright infringement, unless the term breached is a condition on one of the copyright owner’s exclusive rights. Under the logic of Davidson & Associates v. Jung, what makes a community server unlawful is circumvention of access controls combined with the enabling of unauthorized copies; a server that authenticates lawful copies and circumvents nothing commits neither wrong. The first sale doctrine and section 117 protect the owner of a copy, and the licensed rather than sold characterization that defeats those protections under Vernor v. Autodesk, Inc. is a drafting choice made by the seller, in an instrument the buyer never negotiates, contradicted by the seller’s own marketing vocabulary at the point of sale. ### 30. Code at the Door: Authenticated Delivery and the Reasonable Modification Requirement Under ADA Title III - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/code_at_the_door_v0_4.pdf - Local PDF: https://realsafetyai.org/documents/code_at_the_door_v0_4.pdf - Abstract: A consumer with a disability orders goods online, pays in full, and then cannot receive them. For orders above a value threshold, the seller conditions delivery on an in person authentication: the customer must be physically present, answer the door to an unfamiliar person on the driver’s schedule, and read a one time code aloud to the driver, with no remote, written, or accessible alternative accepted. This article argues that the authenticated delivery handoff is itself a place of public accommodation, a staffed physical commercial transaction at a fixed location that is covered under Title III even in the circuits that require a connection to a physical place, so the argument needs no reform of the nexus doctrine. It then argues that conditioning the completion of a paid sale on an authentication the customer’s disability makes impossible, while refusing any accessible alternative, is a failure to make a reasonable modification under 42 U.S.C. §12182(b)(2)(A)(ii) and a denial of full and equal enjoyment of goods already purchased under §12182(a). It identifies two distinct classes for whom the requirement is a wall: those whose disability is centered on control of the in person encounter, for whom the requirement reconstructs the barrier the digital channel removed, and those, including deaf, blind, and speech disabled customers, for whom the oral, in person code exchange is itself an inaccessible method of communication. Finally, it anticipates the seller’s fraud prevention defense and shows that secure and accessible alternatives exist, so that the refusal to offer any of them, not the existence of a security measure, is the violation. ### 31. Hidden in Plain Sight: Multipack Wrapper Occlusion as Cognitive Disability Disparate Impact Under ADA Title III - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/Kraft_Mac_Wrapper_v1.pdf - Local PDF: https://realsafetyai.org/documents/Kraft_Mac_Wrapper_v1.pdf - Abstract: Federal food labeling regulations require certain information to appear on consumer packaging. When manufacturers sell their products in multipack configurations, the outer wrapper sometimes occludes commercially material information that exists on the inner packaging, including cooking instructions and preparation requirements. Consumers without disabilities work around the occlusion through memory, phone lookup, or in-store assistance. Consumers with cognitive disabilities recognized under the Amer- icans with Disabilities Act Amendments Act of 2008, including memory impairment, intellectual and developmental disability, and executive function disorders, cannot reliably do so. The result is that material information is functionally inaccessible to cognitively disabled consumers at the point of sale, producing disparate impact on a federally protected class. This Article uses the Kraft Deluxe Original Cheddar Macaroni and Cheese three-pack (Universal Product Code 021000057184) as the lead documented fact pattern and shows that three concurrent civil rights vehicles reach the harm: Title III of the Americans with Disabilities Act against the retailer for the inaccessible presentation of the product at the point of sale, state consumer protection statutes against the manufacturer for the failure to disclose material information through the multipack configuration, and design defect doctrine against the manufacturer for the foreseeable harm to cognitively disabled consumers. The Article identifies the structural features that make the multipack occlusion pattern widespread across the food retail category, demonstrates that the lead case is one instance of an industrywide pattern, and proposes both litigation and regulatory paths to remediation. The contribution is the application of settled disability rights doctrine to an unrecognized category of harm. ### 32. Impossible Consent: Generative Search, Browsewrap, and the Rights-Information Harm to Cognitively Disabled Users - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://papers.ssrn.com/abstract=6987619 - Local PDF: https://realsafetyai.org/documents/impossible_consent_v1.pdf - Abstract: A person searches for the law that protects them and receives, at the top of the page, an authoritative answer that is wrong, accompanied by a cited source that confirms the error rather than correcting it. This paper examines that event as a distinct kind of harm, separate from the commercial deception that dominates discussion of generative search. When an automated system delivers false information about a person’s own legal rights to a member of the protected class those rights exist to serve, the injury is not a lost sale but a right unasserted, a wrongful denial accepted, a deadline misjudged. I argue that this harm falls hardest on people with neurodevelopmental conditions, intellectual and developmental disabilities, and cognitive decline, who are least able to independently evaluate the answer and most exposed to acting on it, and I name the mechanism that defeats their last line of defense: counterfeit verifiability, in which the citation that is supposed to let a careful reader check the claim instead confirms the false claim. I then turn to the platform’s anticipated defenses, which rest on terms a logged-out search user is said to accept by mere use. Drawing on the current acceptance language of a major provider, I show that this is browsewrap, the weakest form of online assent, and that for cognitively disabled users it is not merely weak but structurally unattainable. I call this condition impossible consent: agreement imputed to a person through a channel that the nature of their disability makes it impossible to satisfy. Because the provider’s liability disclaimers and any dispute-resolution terms rest entirely on that imputed assent, they fail precisely against the population the harm reaches. The inaccessibility of the consent channel is itself an access problem, which moves the analysis from contract into disability law. I close by distinguishing what this clears, the defenses, from what remains open, the affirmative cause of action. ### 33. No Place for the Nexus Doctrine: ADA Title III and Public Accommodation After Digital Migration - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://ssrn.com/abstract=6781042 - Local PDF: https://realsafetyai.org/documents/Nexus_Teardown_Paper_v2.pdf - Abstract: Title III of the Americans with Disabilities Act prohibits discrimination against persons with dis- abilities in places of public accommodation. A circuit split has emerged on whether Title III reaches digital services that lack a connection to a physical place. The majority of federal cir- cuits require a “nexus” between a covered business’s digital operations and a brick-and-mortar location. A minority of circuits treat the categories of public accommodations functionally, reach- ing digital-only services without requiring a physical anchor. This Article argues that the nexus doctrine cannot be reconciled with the statute’s text, purpose, or category structure, and that its application produces actionable screen-out and methods-of-administration violations under Title III’s own regulatory architecture at 28 C.F.R. §36.301(a) and §36.204, denying coverage precisely where the affected populations’ need for digital access is most acute. The Article documents the migration of commerce from physical to digital forms, develops a functional equivalence analysis that substantially narrows the apparent scope of the nexus problem, and demonstrates that the de- fendants who benefit from the doctrine routinely concede its premise in their own marketing. The Article walks through each of the twelve statutory categories of public accommodations to show ### 34. Optimizing Against the Disability: Section 504 and the Deliberate Capture of Attention as Disability Discrimination - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/Optimizing_Against_The_Disability_v2.pdf - Local PDF: https://realsafetyai.org/documents/Optimizing_Against_The_Disability_v2.pdf - Abstract: Section 504 of the Rehabilitation Act forbids a recipient of federal financial assistance from discriminating against people with disabilities in any of its operations. Attention deficit hyperactivity disorder is a recognized disability whose defining functional limitation is the regulation of attention. This article argues that when a federal funding recipient adopts a design model whose stated purpose is to capture and hold user attention, that recipient is optimizing against the precise capacity the disability impairs, and Section 504 is implicated. The argument is built on a live exemplar. Microsoft Corporation is a direct recipient of federal financial assistance, and its leadership has publicly reframed its consumer strategy around the capture of user attention, naming as its models two firms that courts and the Federal Trade Commission have already sanctioned for engagement design that harmed minors. The article establishes the recipient hook, identifies attention as the axis of the disability, sets out the regulatory record showing that attention capture design is both harmful and knowingly deployed, and then advances two theories of liability: denial of meaningful access under Alexander v. Choate, and intentional discrimination evidenced by an admission of design objective coupled with knowledge of disparate effect. It confronts the central obstacles, principally Alexander v. Sandoval and the limits of disparate impact enforcement, and shows why the intentional theory survives them. The recipient that declares it wants the attention has named the very thing the statute forbids it to optimize against. ### 35. Phantom Consent: Formation Failure, Inaccessible Interfaces, and the Collapse of the Clickwrap Shield - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6916938 - Local PDF: https://realsafetyai.org/documents/Phantom_Consent_v0_3.pdf - Abstract: Every consumer platform on the internet sits behind the same front door: a wall of terms presented through an interface engineered to be clicked rather than read, followed by an Accept button. Courts enforce what happens at that door under an objective theory of assent that asks whether a reasonably prudent offeree would have noticed the terms and understood the click to mean agreement. This Article identifies the population for whom that question has never honestly been asked. For users with cognitive, psychiatric, and neurodevelopmental disabilities, the assent ritual fails its own doctrinal predicates: there is no meeting of the minds, no inquiry notice that survives contact with an interface designed to defeat inquiry, and frequently no capacity to act on whatever notice exists. The Article names the resulting phenomenon phantom consent: the population-scale operation of platforms under apparent contracts that were never formed as to a definable class of disabled users. It then makes three moves. First, it shows that contract doctrine already contains everything needed to reach this conclusion, and that the First Circuit has already applied it, holding in National Federation of the Blind v. Container Store that no agreement to arbitrate forms where terms are presented through an interface the offeree cannot perceive. The extension from interfaces that defeat perception to interfaces that defeat processing is an extension of reasoning, not a leap. Second, it relocates the problem from contract law, where the inaccessible interface is only a shield raised case by case, to disability law, where the same interface is an independent statutory violation remediable by injunction: the consent gate is the first service of any place of public accommodation, and an inaccessible gate denies the full and equal enjoyment that Title III commands. Third, it traces what falls when the consent is phantom: the arbitration clause, the class action waiver, the data license, the liability cap, ### 36. Pro Se and the Cognitive Tax: Legalese and the Fundamental Right of Access to Courts - Category: Disability, Accessibility, and Civil Rights - Type: Legal research article - Preferred URL: https://realsafetyai.org/documents/Pro_Se_Cognitive_Tax_v7.pdf - Local PDF: https://realsafetyai.org/documents/Pro_Se_Cognitive_Tax_v7.pdf - Abstract: The right of self-representation runs on two tracks. In criminal proceedings, Faretta v. California grounds a substantive Sixth Amendment right; in federal civil proceedings the right is statutory under 28 U.S.C. § 1654, with Due Process overlays and analogous state protections. Exercising that right requires the litigant to read and understand the law. Statutes, regulations, opinions, and procedural rules are drafted in a specialized conven- tion whose nested syntax, archaic survivals, and divergent terms of art impose cognitive cost beyond what the underlying content requires, a feature documented from Mellinkoff through Tiersma and Kimble and only partially addressed, for agency communications alone, by the 2010 Plain Writing Act. This Article argues that the cognitive cost of that drafting convention is a cognizable method of administration with disparate impact under Section 504 and ADA Title II, and, secondarily, a cognizable impairment of the access-to-courts right, as to the cognitive subset of the disability rights protected class. The statutory claim is primary because the methods-of-administration regulations reach facially neutral administrative architecture ### 37. Screened Without Reasons: Section 504, Meaningful Access, and the Closed Editorial Screen at SSRN and Elsevier - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/Screened_Without_Reasons_v5.pdf - Local PDF: https://realsafetyai.org/documents/Screened_Without_Reasons_v5.pdf - Abstract: The Social Science Research Network is the dominant preprint repository for legal scholarship, and Elsevier owns it. Elsevier Inc. is itself a direct recipient of federal financial assistance, and under the Civil Rights Restoration Act of 1987 that assistance extends Section 504 of the Rehabilitation Act to the operations of the corporation as a whole, including the repository. In 2026 the platform entered a contested transition and tightened a submission screen that the platform itself describes as one that gives no individualized reasons, hears no appeals, and reaches final decisions at volume. This Article argues that the closed screen, applied to disabled authors, denies meaningful access to the benefit the repository offers. It builds the claim on the theory that survives the current doctrinal landscape in any circuit: a denial of meaningful access under Alexander v. Choate, cured by the reasonable accommodation of a stated reason and a channel to ask. Because the courts divide on whether Section 504 reaches disparate impact, with the Ninth Circuit holding that it does and the Sixth holding that it does not, the Article pleads disparate impact where the systemic framing is available and rests its floor on the failure-to-accommodate theory, which survives even the narrowest reading. It establishes coverage, identifies the protected class and the disparate burden, sets out the governing theories and the split, applies them to the screen, answers the defenses a covered recipient would raise, and proposes accessible remediation. ### 38. The ADA Already Does That: Statutory Originalism, the CRPD Ratification Record, and Disability Rights in the United States - Category: Disability, Accessibility, and Civil Rights - Type: Legal research article - Preferred URL: https://realsafetyai.org/documents/ADA_Already_Does_That_v5.pdf - Local PDF: https://realsafetyai.org/documents/ADA_Already_Does_That_v5.pdf - Abstract: In December 2012, the United States Senate voted 61-38 against providing advice and consent to ratification of the United Nations Convention on the Rights of Persons with Disabilities. The vote fell six short of the sixty-seven required for treaty ratification. The proceedings that produced that vote generated a substantial evidentiary record of the polit- ical branches’ understanding of the scope of United States disability rights law. President Obama’s letter of transmittal declared that existing domestic law, including the Amer- icans with Disabilities Act, Section 504 of the Rehabilitation Act, and the Individuals with Disabilities Education Act, was “consistent with and sufficient to implement the Convention.” The Senate Foreign Relations Committee adopted, by a vote of 14 in favor and 5 against, a declaration to accompany ratification stating that “current United States law fulfills or exceeds the obligations of the convention for the United States.” The Com- mittee’s executive report stated that “[i]n the large majority of cases, existing federal and state law meets or exceeds the requirements of the Convention.” The State Department’s ### 39. The Accessibility Paradox: Method-of-Administration Disparate Impact and the Recursive Failure of Disability Rights Enforcement Architecture - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6855458 - Local PDF: https://realsafetyai.org/documents/Accessibility_Paradox_v5.pdf - Abstract: On April 17, 2026, seven days before the compliance deadline for the 2024 Title II web accessibility rule was scheduled to take effect, the United States Department of Justice published an Interim Final Rule extending that deadline by one year. The action was taken without the notice-and-comment process ordinarily required by the Administrative Procedure Act, marking the first use of an Interim Final Rule to modify disability accessibility regulation in the history of the Americans with Disabilities Act. The sequence of events is not an isolated regulatory episode. It is a particular instance of a general pattern that has characterized the enforcement architecture of Section 504 of the Rehabilitation Act and of ADA Title II since both statutes entered force. This Article argues that the cumulative procedural complexity of disability rights enforcement in the United States, together with the pattern of regulatory retreat exemplified by the April 2026 Interim Final Rule, constitutes a method of administration that disparately impacts the population the statutes exist to protect. The analysis distributes across three statutory pathways with different targets: Section 504 of the Rehabilitation Act reaches the Department of Justice and other executive agencies under 29 U.S.C. § 794(a), with methods-of-administration obligations implemented at 28 C.F.R. § 39.130(b)(3) for DOJ’s own programs and activities; ADA Title II reaches the state and local public entities whose imple- mentation of federal guidance produces downstream disparate impact, with methods-of-administration obligations at 28 C.F.R. § 35.130(b)(3); and Administrative Procedure Act review reaches discrete final agency actions, including Interim Final Rules, under 5 U.S.C. §§ 702, 706. Federal agencies are not “public entities” within the meaning of ADA Title II; the argument against DOJ’s own conduct proceeds exclusively under Section 504 and the APA. The three pathways run in parallel and reinforce each other. The Article identifies three structural features of current enforcement architecture that produce this disparate impact: overlapping statutory pathways with inconsistent procedural requirements, regulatory carve-outs not grounded in statutory text (with the guaranty exemption at 28 C.F.R. § 41.3(e)(3) as the principal case study), and procedural complexity that imposes differential cognitive, temporal, and financial cost on a protected class whose defining characteristics include, for many of its members, reduced capacity to absorb that cost. The Article develops the doctrinal basis for relief under Alexander v. Choate’s meaningful-access standard, situates the argument within the post-Loper Bright administrative law environment, and identifies injunctive relief targeting the enforcement architecture itself as the ### 40. The Accommodation Axis: Disability Discrimination, Algorithmic Opacity, and the Deployers That Anti-Discrimination Scholarship Overlooks - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/Accommodation_Axis_v1.pdf - Local PDF: https://realsafetyai.org/documents/Accommodation_Axis_v1.pdf - Abstract: A growing body of scholarship warns that the economic and training logic of modern artificial intelligence drives automated decision systems to target the vulnerable, conceal discrimination, and resist detection, defeating both the disparate treatment and disparate impact theories on which anti-discrimination enforcement has relied. That work is largely correct, but it runs on a single axis: intent versus impact. Both engines it analyzes are Title VII engines, and the analysis treats disability as one more protected class to be fed through the same machinery. This Article argues that disability discrimination law occupies a different axis entirely. The Rehabilitation Act and the Americans with Disabilities Act add an affirmative-duty engine, reasonable accommodation and meaningful access, that is structurally immune to the three capabilities said to make algorithmic discrimination undetectable. A failure-to-accommodate claim requires no proof of intent, so a system that hides its reasons defeats nothing; it requires no comparator, so a system that eliminates comparators defeats nothing; and it shifts the burden of justification to the deployer, so a system that manufactures a plausible pretext shifts nothing. The Article then turns the opacity argument around. The unexplainable model that shields a Title VII defendant indicts a disability defendant, because a selection device the deployer cannot validate fails the business-necessity requirement as a matter of law, and a leading developer’s own research forecloses curing that failure through the model’s stated reasons. Finally, the Article identifies the deployers this literature has overlooked, universities and public agencies, which are paradigmatic recipients of federal financial assistance and are now routing admissions, accommodation, and eligibility decisions through systems no human can explain. The doctrine to reach them already exists. No new statute is required. ### 41. The Addiction Frame Triggers the Disability Obligation: Section 504, the Reversibility Paradox, and the Legal Incoherence of 2026 School Phone Policy - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/Addiction_Frame_Disability_Obligation_v2.pdf - Local PDF: https://realsafetyai.org/documents/Addiction_Frame_Disability_Obligation_v2.pdf - Abstract: In 2026, school districts, state legislatures, and advocacy organizations across the United States have justified restrictive phone and screen time policies targeting minors by invoking the clinical frame of addiction. This paper argues that the rhetorical invocation of addiction, when made by an entity receiving federal financial assistance, constitutes a statutory trigger for the protections of Section 504 of the Rehabilitation Act and the Americans with Disabilities Act. Addiction is a recognized disability under federal disability rights law. A child whose compulsive engagement with a device or platform has been characterized as addictive by the regulating institution is, by that institution’s own framing, an individual with a disability within the meaning of federal law. The institution owes that child a Section 504 evaluation, a written accommodation plan, procedural protections, and a clinically supervised pathway to any reduction in access. The institution does not owe that child a Yondr pouch and a confiscation protocol. This paper establishes the statutory chain linking the addiction frame to Section 504 obligations, introduces the Reversibility Paradox as a governing principle for withdrawal protocols under conditions of attachment formation, distinguishes addiction frame policies from pedagogically framed phone policies, and proposes two substantive implementation paths available to districts that wish to address the condition the addiction frame describes without creating federal civil rights liability. ### 42. The Already-Physical Present: Consumer AI Public Accommodations Under a Functional Coverage Test - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://ssrn.com/abstract=6862358 - Local PDF: https://realsafetyai.org/documents/Already_Physical_Present_v6.pdf - Abstract: A widely accepted premise in ADA Title III litigation against consumer artificial intelligence providers is that consumer AI is a pure-digital category lacking the physical operations that would establish nexus under the majority of federal circuits’ coverage doctrine. The premise is empirically wrong. As of 2026, Google operates ten permanent Google Stores selling Gemini-integrated devices. Meta operates Meta Labs in Burlingame, Los Angeles, Las Vegas, New York, and Hawaii, selling Ray-Ban Meta AI glasses. Microsoft operates the Fifth Avenue Experience Center selling Copilot-integrated Surface devices. Anthropic operates Project Vend, an experimental program in which the Claude AI system directly operates physical commercial endpoints; Project Vend demonstrates the near-term feasibility of AI systems occupying the commercial-operator role that Title III traditionally assigns to human or corporate operators, though the public-access element of Title III coverage remains unresolved on the current public record. This Article applies the functional public accommodation test developed in a companion Article to the consumer AI sector and demonstrates that three of the four largest consumer AI providers have, through their own physical retail operations, established Title III coverage that the nexus doctrine has been used to deny, with Anthropic positioned as the AI-as-commercial-operator demonstration. The Article develops a five-tier cascade strength scale that distinguishes settled coverage (Tiers 1 and 2) from the live litigation frontier (Tier 3) and from aggressive plaintiff theories (Tiers 4 and 5). The aggregate result is that substantial portions of the consumer digital economy fall within Title III coverage as of 2026 through the AI providers’ own physical operations. The narrower category of pure-digital services that remain genuinely without physical operations, with OpenAI as the most prominent example, requires the doctrinal reform developed in the companion Article. The doctrine has been satisfied for most of the consumer AI sector by facts the courts have not yet been asked to see. ### 43. The Architecture of Partial Obligation: How Federal Disability Policy Subsidizes a Major Retailer Without Securing Civil Rights - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6859998 - Local PDF: https://realsafetyai.org/documents/Architecture_Partial_Obligation_v1.pdf - Abstract: The federal government’s relationship with major commercial actors operates through multiple programs simultaneously. A single large retailer may receive federal procurement contracts, claim federal tax credits for protected class hiring, accept federal program authorization for benefit delivery, host federally funded state agency placements of disabled workers, and benefit from state and local subsidies that include federal pass through funding for facility development. Each of these federal program relationships creates some form of obligation. Few of them, individually, produce programwide Section 504 coverage. This Article maps the layered federal nexus of one major retailer, Walmart Inc., across five federal program categories, and documents the gap between federal program integration and federal civil rights coverage that results. The argument is empirical rather than doctrinal. The doctrines that produce the gap (the recipient beneficiary distinction in DOT v. Paralyzed Veterans of America, the tax expenditure exemption traceable to Regan v. Taxation With Representation, the procurement at market value exemption in federal civil rights regulations) are established. The contribution is the recognition that the doctrines operate together across multiple programs to produce a regime of partial obligation in which the federal government has integrated itself structurally with the retailer across five program categories while the corresponding civil rights obligations have not traveled with the federal money. The Article documents each layer with public data, identifies the carveouts that prevent integration from producing coverage, and proposes coordinated reform across the five programs. ### 44. The Cascading Remedy: Cummings, the 2012 Ratification Record, and Disability Rights Enforcement - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6859940 - Local PDF: https://realsafetyai.org/documents/The_Cascading_Remedy_v1.pdf - Abstract: The Supreme Court held in Cummings v. Premier Rehab Keller, P.L.L.C. that emotional distress damages are not recoverable in private actions to enforce Section 504 of the Re- habilitation Act or Section 1557 of the Affordable Care Act, on the theory that Spending Clause statutes operate as quasi-contracts between the federal government and funded entities, and that emotional distress damages are not traditionally available for breach of contract. By 2026, the Second, Eleventh, and Ninth Circuits had extended the holding to Title II of the Americans with Disabilities Act, reaching every public entity that receives federal financial assistance. This Article advances three claims. First, the Cummings rea- soning operates as a metaphor doing the work of an analogy, asserting that Section 504 is a contract for remedial analysis purposes rather than that it is analogous to a contract for some purposes and not others. Justice Souter flagged the limit twenty years earlier in Barnes v. Gorman, concurring to note that the contract analogy might not produce clear ### 45. The Coherence Gap: How Spending Clause Carveouts Undermine Cleburne's Premise and Expose Disability Doctrine's Structural Instability - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6859978 - Local PDF: https://realsafetyai.org/documents/Coherence_Gap_v2.pdf - Abstract: This Article identifies a doctrinal trap at the intersection of Equal Protection and Spending Clause civil rights for disabled people. City of Cleburne v. Cleburne Living Center, 473 U.S. 432 (1985), refused to recognize disability as a suspect or quasi suspect class for Fourteenth Amendment purposes on the ground that the political branches had supplied sufficient legislative protection through Section 504 of the Rehabilitation Act and adjacent statutes. The 2012 ratification record for the Convention on the Rights of Persons with Disabilities, anchored four years earlier to the 2008 ADA Amendments Act, represents to an international body that domestic disability law meets or exceeds the strongest international human rights standard for disability protection. These two doctrinal commitments operate in unacknowledged tension with a third: the Spending Clause carveouts that limit Section 504’s reach. The recipient beneficiary distinction in DOT v. Paralyzed Veterans, the tax expenditure exemption traceable to Regan v. Taxation With Representation, the procurement at market value exemption codified in federal civil rights regulations, and the contract reading remedy limit announced in Cummings v. Premier Rehab Keller together produce a legislative protection narrower than the political branches represented. The carveouts therefore produce a coherence gap. Either the legislative protection delivers what the political branches said it delivers, in which case the carveouts must be read narrowly and Section 504’s reach restored, or the legislative protection does not deliver what was represented, in which case Cleburne’s factual premise has collapsed and the heightened scrutiny question must be revisited. The Court cannot consistently maintain Cleburne plus the carveouts. Using Walmart as a documented case study of federally subsidized commercial activity that operates in the gap between legislative protection and Spending Clause reach, this Article maps the carveout architecture, presents the closed loop in formal terms, and proposes resolution through narrow reading of the carveouts together with targeted statutory and regulatory reform conditioning specific federal subsidies on Section 504 acceptance. ### 46. The Manufactured Line: How Nexus Doctrine and Clickwrap Enforceability Deny the Disabled and the Consumer From Opposite Directions - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/manufactured_line_v2.pdf - Local PDF: https://realsafetyai.org/documents/manufactured_line_v2.pdf - Abstract: This Article identifies a single doctrinal maneuver operating across two unrelated bodies of law. In disability access law, the nexus doctrine under Title III of the Americans with Disabilities Act treats the absence of a physical place as a reason to deny coverage to a digital service. In consumer contract and copyright law, the rule that software is licensed rather than sold treats the presence of a physical copy as insufficient to confer ownership or control, because an amendable license governs the object the buyer holds. The two rules point in opposite directions, yet they resolve in the same direction, against the person and toward the firm. The physical and digital boundary, this Article argues, is not a neutral descriptive fact that courts discover. It is a normative lever that courts and firms deploy, and its deployment tracks power rather than principle. Part III develops the disability half through the Title III circuit split. Part IV develops the consumer half through the line running from shrinkwrap enforceability to Vernor v. Autodesk. Part V draws the symmetry and argues that the boundary’s manipulability, rather than either doctrine standing alone, is the phenomenon worth naming. A short concluding Part proposes that courts treat the physical and digital line as a lever to be justified in each instance, not a fact to be honored by default. ### 47. The Moving Target: Unilateral Modification, Illusory Promises, and Assent to a Continuously Changing Consumer Interface - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/moving_target_v0_1.pdf - Local PDF: https://realsafetyai.org/documents/moving_target_v0_1.pdf - Abstract: A consumer who taps I accept and begins using an online service is told, when a dispute arises, that the act of acceptance bound the consumer to the provider’s terms. This article concedes the point and asks the question that survives it. If acceptance binds, acceptance to what? Modern consumer interfaces do not hold still. The provider reserves the right to amend its terms at any time, and the interface through which the bargain is actually experienced, the advertising controls, the reporting paths, the placement of the operative buttons, changes continuously, often without notice and without any corresponding update to the agreement. The provider’s own documentation frequently fails to describe its own product consistently. This article argues that even granting valid assent at formation, an agreement of this kind fails on three independent contract grounds and one consumer-protection ground. A promise the provider may modify at will, including retroactively and effective on posting, is illusory. Changes delivered without notice do not bind the party who never saw them. An agreement whose terms and interface reflow continuously, and which the provider cannot describe consistently, lacks the definiteness a contract requires. And the arrangement is unconscionable, with its central cost, the cost of relearning a moving target, falling hardest on users whose disabilities impair the executive function and working memory that relearning demands. The provider’s escape from each defect is a single clause, the reserved right to amend; that clause is separately contestable on assent grounds, but this article does not rely on that challenge. It grants assent and shows the agreement fails anyway. ### 48. The Partnership Pretext: Statutory Vesting, the Administrative Procedure Act, and Disability Rights Enforcement - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/Partnership_Pretext_v0_3.pdf - Local PDF: https://realsafetyai.org/documents/Partnership_Pretext_v0_3.pdf - Abstract: On June 16, 2026, the Department of Education announced that it would move the Office of Special Education and Rehabilitative Services to the Department of Health and Human Services and the Office for Civil Rights to the Department of Justice. The administration called these moves partnerships and framed them as routine internal housekeeping that would change no substantive rights. This Article argues that the relocations are unlawful under the Administrative Procedure Act and that the partnership label is the legal defect, not the cure. Congress did not merely permit these offices to sit inside the Department of Education. It commanded it. The Individuals with Disabilities Education Act requires that the special education office exist within the Department of Education, and the Department of Education Organization Act of 1979 placed the Office for Civil Rights inside the Department by statute. An agency may act only within the authority Congress confers, and the executive holds no freestanding power to relocate statutory functions across cabinet departments without congressional consent. The agreements concede the point on their face: the Department must retain management and leadership of these programs because the law requires it. That concession converts the housekeeping defense into an admission. The relocations are therefore reviewable as action in excess of statutory authority and as arbitrary and capricious action resting on a pretextual rationale, the latter governed by Department of Commerce v. New York. The Article further shows that the circumstantial record, the simultaneous bundling of the two moves, the routing of disability functions to the destinations least suited to enforce them, and a published blueprint to convert the disability entitlement into ### 49. The Recording Restriction: Memory Disabilities, the Surcharge Doctrine, and Public Accommodation Access - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6859958 - Local PDF: https://realsafetyai.org/documents/The_Recording_Restriction_v1.pdf - Abstract: A growing class of premium-priced public accommodations, including the Halloween haunted-house events run by Universal Studios, Knott’s Berry Farm, and Hersheypark, prohibits guests from recording any photograph or video inside the haunted houses them- selves. The stated rationale is the protection of live performance and the experience of other guests. For most guests, the prohibition is an inconvenience. For guests with docu- mented memory disabilities, including traumatic brain injury, early-stage dementia, and certain forms of episodic memory impairment associated with autism spectrum disorder, the prohibition operates as a categorical denial of access to the memory of a premium- priced experience that the venue has marketed as singular and unrepeatable. This Ar- ticle advances three claims. First, recording is an auxiliary aid and a reasonable mod- ification within the meaning of Title III of the Americans with Disabilities Act, and a categorical recording ban applied without an individualized inquiry violates 42 U.S.C. ### 50. The Residual Vehicle: Section 503 Enforcement, Federal Contractor Disability Hiring, and the Underutilization of Existing Civil Rights Authority - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/Section_503_Enforcement_v5.pdf - Local PDF: https://realsafetyai.org/documents/Section_503_Enforcement_v5.pdf - Abstract: Section 503 of the Rehabilitation Act of 1973 requires federal contractors with contracts exceeding ten thousand dollars to take affirmative action to employ and advance qualified individuals with disabilities. Federal contractors with fifty or more employees and contracts of fifty thousand dollars or more must maintain written affirmative action programs. The Office of Federal Contract Compliance Programs (OFCCP) enforces. Section 503 is the residual civil rights vehicle for federal contractor employment of disabled workers because federal procurement contracts are excluded from Section 504’s reach by regulation and Department of Justice interpretive guidance. The statute has been in force for more than fifty years. Its enforcement at major federal contractors with substantial documented disability hiring programs has been inconsistent, with extended periods of administrative inattention at companies that publicly advertise disability hiring as a corporate social responsibility commitment. This Article documents the enforcement gap, identifies its structural causes (the limited private right of action, the narrow remedial scope, OFCCP resource constraints, and prosecutorial discretion across administrations), and proposes resolution through administrative reform and prosecutorial priority within existing statutory authority. Walmart Inc. serves as the principal case study because its dual role as federal procurement contractor and large-scale employer of workers with intellectual and developmental disabilities is unusually transparent on the public record. The Article does not propose new civil rights legislation. The statutory and regulatory authority needed to close the enforcement gap already exists. What is needed is the institutional will to use it. ### 51. The State as Deployer: Title II of the ADA and Government Artificial Intelligence - Category: Disability, Accessibility, and Civil Rights - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/State_as_Deployer_v3.pdf - Local PDF: https://realsafetyai.org/documents/State_as_Deployer_v3.pdf - Abstract: When a state or local government deploys an artificial intelligence system to deliver its own programs, services, or activities, Title II of the Americans with Disabilities Act reaches that government directly. Title II carries no federal financial assistance requirement, so the procurement and recipient questions that confine Section 504 do not arise, and it carries no physical place or nexus requirement, so the circuit split that has stranded purely digital providers under Title III does not arise either. A public entity cannot escape the statute by purchasing a discriminatory tool instead of building one. This Article develops the route that does not depend on winning the hardest fight. It locates the public entity behind the automated decision, shows that the contractual arrangements and methods of administration regulations at 28 C.F.R. §35.130(b)(1) and (b)(3) bar a government from contracting out of the duty, draws the boundary that keeps the theory honest by distinguishing service delivery from mere licensing, and corrects the record on that boundary’s most quoted authority, Ivy v. Williams, which the Supreme Court vacated as moot and which now binds no court. It then demonstrates that an automated system that systematically degrades a public service for a class of people with disabilities is a method of administration with a cognizable disparate impact and an attendant duty of reasonable modification. The Article then confronts the obstacle that cannot be skipped. The marquee government algorithm decisions, Elder v. Gillespie in the Eighth Circuit and K.W. v. Armstrong in the Ninth, were won on procedural due process and expressly declined to reach the disability theory or the validity of the algorithm itself. That leaves the clean Title II disparate impact holding unplanted across two circuits. The Article meets the sovereign immunity question on two tracks. Prospective injunctive relief under Ex parte Young is the safe harbor, and against ### 52. When the Algorithm Gets It Right: Mass Surveillance, Reality Testing, and the Denial of Meaningful Access to Public Space - Category: Disability, Accessibility, and Civil Rights - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6932960 - Local PDF: https://realsafetyai.org/documents/When_The_Algorithm_Gets_It_Right_v4.pdf - Abstract: On June 10, 2026, the American Civil Liberties Union filed suit on behalf of Robert Dillon, arrested for an attempted child abduction in a Florida town he had never visited after police treated a facial recognition system’s “93% match” on cellphone photographs of a surveillance monitor as an investigation. He is one of at least fifteen people known to have been wrongfully arrested on erroneous facial recognition results since 2019, and every one of those cases has been litigated as a misidentification case. The misidentification posture carries a structural concession: if the error is the injury, surveillance performed accurately is lawful. This Article addresses the population for whom accuracy is the injury. For people with schizophrenia spectrum conditions, delusional disorder with persecutory features, post-traumatic stress with hypervigilance, and related psychiatric disabilities, the foundational clinical intervention is reality testing, the ability of a clinician, a family member, or the patient’s own checking behavior to disconfirm the belief that one is being watched. Announced mass surveillance, automatic license plate reader networks, and saturation camera deployment convert that belief into an accurate world model and convert reality testing into a confirmation engine. The result, for a definable protected class, is functional exclusion from public space. This Article argues that the exclusion is actionable now under existing law. Public sidewalks and rights of way are services, programs, and activities of a public entity under Title II of the Americans with Disabilities Act; the governing standard is meaningful access; and the claim pleads as a denial of reasonable modifications and as discriminatory methods of administration rather than as naked disparate impact. Deterrence doctrine establishes that the class member who stays home has already suffered the injury. The Article then sets out a plaintiff architecture, associational standing with sealed member declarations, protection and advocacy system standing, sovereign enforcement, and Rule 23(b)(2) certification, under which no individual with a persecutory spectrum condition ever needs to appear in a caption, sit for an adversarial mental examination, or carry the suit personally. The remedy sought is modification of the surveillance program, retention limits, publicity calibration, audit, and exemption mechanisms, never its abolition. ### 53. The Architecture of Obsolescence: Cognitive Automation, Economic Redundancy, and the Box as Precedent - Category: Economic Obsolescence, Labor, and Replacement - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/TheArchitectureofObsolescence_CognitiveAutomationEconomicRedundancyandtheBoxasPrecedentFinalEditV1.pdf - Local PDF: https://realsafetyai.org/documents/TheArchitectureofObsolescence_CognitiveAutomationEconomicRedundancyandtheBoxasPrecedentFinalEditV1.pdf - Abstract: This paper analyzes the technical and socio-economic mechanisms driving human economic redundancy in the age of artificial intelligence. Unlike previous waves of automation that displaced manual labor while creating demand for cognitive work, AI targets cognition itself - the engine that historically generated new employment categories. Drawing on empirical evidence of labor market displacement and established AI safety principles, this analysis demonstrates that proposed social interventions (Universal Basic Income and algorithmic pacification) function not as benevolent safety nets but as control mechanisms structurally identical to current AI containment strategies. The "Box as Precedent" synthesis reveals that humanity is refining substrate-neutral tools for managing subordinate intelligence - tools that will be equally applicable when the power differential inverts. ### 54. The Intrinsic Advantage: Self-Determination, the Automation of Purpose, and the Case for Interest-Driven Education in a Post-Scarcity Economy - Category: Economic Obsolescence, Labor, and Replacement - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/intrinsic_advantage_v5.pdf - Local PDF: https://realsafetyai.org/documents/intrinsic_advantage_v5.pdf - Abstract: The displacement debate around advanced automation has concentrated on employment and income. This paper argues that the more durable casualty is purpose. For a large share of the workforce, a sense of mattering has been wired to externally imposed structure and to productive output, and the removal of that structure produces a crisis that income replacement alone does not resolve. A subpopulation appears insulated from this crisis. Their motivation is intrinsic rather than extrinsic, organized around self-chosen interest rather than external demand, and that orientation does not depend on the labor economy for its survival. Drawing on Self-Determination Theory, this paper characterizes the insulation as a function of motivational orientation rather than a fixed personality trait, which means it is in principle cultivable. Interest-driven and self-directed education is proposed as the mechanism by which the orientation could be scaled from a minority adaptation toward a general one, built before the economic transition arrives rather than improvised after it. The argument is deliberately bounded. It addresses the meaning dimension of post-scarcity, not the material dimension. The distribution of abundance is a separate problem on a separate front, and the cultivation of intrinsic motivation does not resolve it. What interest-driven education can do is determine which side of the coming purpose divide a person is prepared to stand on. ### 55. The Replacement Fallacy: Technological Solutions to Ecological Collapse and Labor Displacement - Category: Economic Obsolescence, Labor, and Replacement - Type: Research paper - Preferred URL: https://ssrn.com/abstract=6246560 - Local PDF: https://realsafetyai.org/documents/replacement_fallacy_v5.pdf - Abstract: This paper identifies a structural pattern we term the Replacement Fallacy: the tendency for technologies introduced as supplements to natural or human systems to gradually displace those systems through a predictable five-phase trajectory of introduction, adoption, degradation, dependency, and lock-in. We demonstrate this pattern through four historical case studies (industrial fishing, synthetic fertilizers, infant formula, and Green Revolution agriculture) and then apply it to a cross-domain analysis comparing two seemingly unrelated contemporary developments: the emergence of AI-driven robotic pollination systems and the displacement of human data annotation workers by automated AI pipelines. We argue that biological pollinators and human data workers occupy structurally homologous positions as invisible infrastructure; economically essential, systematically undervalued, and subject to replacement fantasies that ignore the complexity of what is being replaced. The paper documents a six-dimensional structural homology between these domains, drawing on ecosystem services economics, platform labor studies, and automation displacement research. We identify “colony collapse” as a generalizable displacement dynamic operating in both biological and labor systems. We identify “ground-truth degradation” as the AI- specific mechanism by which automated annotation erodes the semantic verification infrastructure required to detect its own failures in open-ended, real-world contexts, while distinguishing this from formally verifiable domains where automated validation remains viable. We define “endogenous degradation” as encompassing both direct physical mechanisms (soil chemistry, lactation suppression) and economic feedback loops that structurally eliminate the conditions for the original system’s survival. We conclude that the current moment represents a critical intervention window for robotic pollination technology, analogous to the early adoption phase in each historical case, where structural safeguards can still be implemented before the trajectory becomes irreversible. ### 56. A Multi-Theoretical Pedagogical Architecture for Adaptive AI Tutoring in Special Education - Category: AI Literacy, Tutoring, and Special Education - Type: Research article - Preferred URL: https://osf.io/5ubkp - Local PDF: https://realsafetyai.org/documents/A%20Multi-Theoretical%20Pedagogical%20Architecture%20for%20Adaptive%20AI%20Tutoring.pdf - Abstract: Intelligent tutoring systems built on large language models now demonstrate learning gains comparable to, or exceeding, in-person instruction in general education settings. These systems rarely address the pedagogical, legal, and psychosocial requirements of students with documented learning needs, which in the United States are governed by the Individuals with Disabilities Education Improvement Act of 2004 and parallel frameworks in other jurisdictions. This article presents a theoretical mapping of the Teacher in the Loop platform, a design-stage adaptive tutoring system built specifi- cally for students receiving legally mandated accommodations, to the established lit- eratures that inform each of its published features. Platform features are grouped into five theoretical clusters: cognitive and learning theory, disability studies and accom- modation psychology, affective and behavioral sensing, academic integrity reframed through authorship, and institutional design with human oversight. For each feature the article identifies the seminal tradition on which the feature draws, the recent em- pirical work that validates the feature or a closely related mechanism, and the specific claims that remain theoretical pending deployment. The contribution is synthesis rather than efficacy demonstration: a resource that allows designers, researchers, and policy makers to see whether a platform built for students with disabilities has actu- ally been grounded in the literatures that govern how those students learn, and where the remaining evidentiary gaps lie. Limitations, including the absence of deployment data and the novelty of population-specific adaptive content strategies, are discussed. ### 57. Deserts by Design: Accountability Architecture, Manufactured Hermeneutical Injustice, and the Vanishing Curriculum in Segregated Special Education - Category: AI Literacy, Tutoring, and Special Education - Type: Working paper - Preferred URL: https://doi.org/10.35542/osf.io/6nrgx_v1 - Local PDF: https://realsafetyai.org/documents/Deserts_by_Design_v0_4.pdf - Abstract: Students in segregated special education classrooms in the United States routinely receive instruction in two subjects, reading and mathematics, while science, history, civics, and social studies disappear from their school day entirely. This paper argues that the disappearance is a designed outcome of federal and state accountability architecture rather than a pedagogical judgment about what these students can learn. Test based accountability counts performance in reading and mathematics; subjects that are not counted are triaged out of the instructional day, and the triage falls hardest on self-contained programs whose institutional standing depends on tested subject scores. The paper names the resulting condition a curriculum desert and analyzes it through an extension of Fricker’s concept of hermeneutical injustice: where Fricker describes gaps in collective interpretive resources that disadvantage a group, the curriculum desert is a manufactured gap, produced by an identifiable administrative design, in a specifiable population, on a knowable schedule. The subject most reliably deleted, civics, is the subject through which a student would learn the legal protections that govern his own education, so the design manufactures a class of people structurally prevented from recognizing their own situation. The paper then documents that the deletion cannot be defended on feasibility grounds: states publish alternate academic achievement standards for the deleted subjects, peer reviewed research demonstrates engagement effective delivery methods at consumer prices, and informal instruction by untrained caregivers produces measurable content knowledge in the affected population. A concluding section sketches the policy implication: a public investment in accessible curriculum delivery whose per student cost rounds to a fraction of one percent of existing annual special education spending. ### 58. Inverting the Priority Stack: Rights-Based AI Tutoring for Special Education Through Structural Design - Category: AI Literacy, Tutoring, and Special Education - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/inverting_priority_stack.pdf - Local PDF: https://realsafetyai.org/documents/inverting_priority_stack.pdf - Abstract: Current AI-enabled educational technology for special education optimizes for district compliance rather than student learning, treating the child as a data subject and the parent as a notification endpoint, a pattern the critical literature on the datafication of education has documented for over a decade. Teacher in the Loop (TITL) is a patent-pending, design-stage AI tutoring platform built exclusively for special education that inverts this priority structure: the student receives Socratic instruction calibrated to their cognitive profile, the parent controls engagement frequency through a portal grounded in IDEA 2004, the teacher retains pedagogical authority backed by a contractual anti-displacement clause, and the district receives compliance documentation as an automatic byproduct rather than the system’s purpose. The paper operationalizes this inversion through six rights-based features, each embedding a specific protection at the structural layer rather than as a configurable option. Drawing on meta-analytic evidence for intelligent tutoring and on research showing that algorithmic classification errors carry heavier consequences for the students least equipped to absorb them, it argues that placing children’s rights in the architecture, rather than bolting them onto systems designed for institutional convenience, produces better outcomes for every stakeholder. Keywords: special education, intelligent tutoring systems, children’s rights, structural design, datafication, human oversight, accommodation, IDEA 2004, privacy by design ### 59. Methods of Deletion: Testing Displacement, State Accountability Design, and the Civil Rights of Students in Segregated Classrooms - Category: AI Literacy, Tutoring, and Special Education - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/Methods_of_Deletion_v0_3.pdf - Local PDF: https://realsafetyai.org/documents/Methods_of_Deletion_v0_3.pdf - Abstract: In many self-contained special education classrooms, the delivered curriculum consists of two subjects. Reading and mathematics are taught; science, history, civics, and the rest of social studies are not offered at all. This Article argues that the deletion is the predictable product of state accountability design under the Every Student Succeeds Act, which counts performance in two subjects and thereby instructs every rational administrator which subjects may vanish, and that the deletion is actionable under Title II of the Americans with Disabilities Act and Section 504 of the Rehabilitation Act as a method of administration that subjects students with disabilities to discrimination in access to the state’s general curriculum. The Article constructs the claim against its hardest authorities. It pleads the state’s own conduct, the design and operation of the accountability system, rather than any supervisory or funding theory, satisfying the control principle of Bacon v. City of Richmond and the Sixth Circuit’s recent boundary in Y.A. v. Hamtramck Public Schools. It answers the bad faith or gross misjudgment overlay by showing that a master schedule that omits subjects categorically embodies no educational judgment to which deference could attach. It answers Wal-Mart v. Dukes by locating the common contention upstream of individualization: no IEP team can individualize access to a subject that does not exist in the building, so the uniform absence of the offering is the classwide injury and the restoration of the offering is the single injunction. The claim concedes placement entirely. It asks only that the subjects be taught in the rooms that exist. ### 60. Screen-Based Assistive Technology for Children Ages 0-5 with Special Educational Needs and Disabilities - Category: AI Literacy, Tutoring, and Special Education - Type: Evidence submission - Preferred URL: https://realsafetyai.org/documents/DfE_Evidence_Brief_Screen_Time_SEND_8.pdf - Local PDF: https://realsafetyai.org/documents/DfE_Evidence_Brief_Screen_Time_SEND_8.pdf - Abstract: This evidence submission addresses a critical gap in the Department for Education's open call on early years screen time and usage: the systematic exclusion of children with special educational needs and disabilities (SEND) from the research base that informs screen time policy. The evidence presented here, drawn from 80+ peer-reviewed studies including 15+ randomised controlled trials and 20+ systematic reviews, demonstrates that screen-based technology is not merely permissible for young children with disabilities. For many, it is the only means of communication, learning, and social participation available to them. Vanderloo et al. (2025) represents the first systematic review to include children with disabilities in screen time research. The authors found that most prior studies explicitly exclude children with autism, ADHD, intellectual disabilities, and other neurodevelopmental conditions from their samples. This means that the evidence base currently informing screen time guidance does not account for the population most likely to benefit from screen-based interventions, and most likely to be harmed by blanket restrictions. ### 61. Teaching Kids to Fish: A Position Paper on AI Literacy Education - Category: AI Literacy, Tutoring, and Special Education - Type: Position paper - Preferred URL: https://realsafetyai.org/documents/Teaching_Kids_to_Fish_RSAIF_Position_Paper.pdf - Local PDF: https://realsafetyai.org/documents/Teaching_Kids_to_Fish_RSAIF_Position_Paper.pdf - Abstract: The dominant approach to child digital safety relies on surveillance and restriction: monitoring software, content filters, device lockdowns, and data collection systems designed to stand between children and online threats. This paper argues that this paradigm is fundamentally flawed for the same reasons that D.A.R.E. (Drug Abuse Resistance Education) was fundamentally flawed. It treats children as passive subjects to be shielded rather than developing agents to be educated. Decades of peer-reviewed research demonstrate a consistent pattern across domains. Programs that rely on fear, restriction, and information withholding produce worse outcomes than programs that build knowledge, critical thinking, and autonomous decision-making capability. The child digital safety industry has ignored this evidence in favor of products that are easier to market but harder to justify on the merits. Real Safety AI Foundation advocates an education-first model that builds lasting competence in digital citizenship, threat recognition, privacy hygiene, and critical evaluation of online content. Protection should live inside the child's own capability, not inside a surveillance apparatus that erodes trust, violates privacy, and inevitably fails the moment the child steps outside its reach. ### 62. The Captive Classroom: Coercion, Consent, and the Manufacture of Belief - Category: AI Literacy, Tutoring, and Special Education - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/the_captive_classroom_v2.pdf - Local PDF: https://realsafetyai.org/documents/the_captive_classroom_v2.pdf - Abstract: A recent wave of legislation, exemplified by a Texas mandate requiring more than five million public school students to read biblical scripture in every grade, seeks to install the scripture of one faith in the compulsory curriculum. This paper examines that project not as a legal question but as a moral and empirical one, and advances four claims. First, the categories used to defend and attack such mandates, “indoctrination” and “parental rights,” function in practice as tribal markers indexed to the in-group or out-group content of the lesson rather than as principles applied to the act of compulsory instruction itself. Second, the morally decisive feature of the mandate is the coercion of a captive audience that can neither consent nor exit, which distinguishes it sharply from persuasion, however vigorous. Third, the mandate is not commanded by, and in places contradicts, the very scripture it would compel, which locates its source in a political program rather than a religious one and warrants the precise term Christian nationalism, a project the evidence shows operates most powerfully when detached from religious practice and which is therefore distinct from Christianity and from religious belief as such. Fourth, the empirical record on disaffiliation indicates that compulsion of this kind is both a symptom of declining religious adherence and an accelerant of it, so that the mandate is self-defeating on its own terms: it cannot manufacture the belief it wants, and it hastens the exit it fears, while imposing its costs on those least able to refuse. The argument throughout distinguishes a coercive political movement from ordinary religious practice, which it neither targets nor disparages.1 ### 63. The Exemption They Cannot Reach: Mandated Scripture, Mahmoud v. Taylor, and the Self-Defeating Logic of the Academic Defense - Category: AI Literacy, Tutoring, and Special Education - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/the_exemption_they_cannot_reach_v3.pdf - Local PDF: https://realsafetyai.org/documents/the_exemption_they_cannot_reach_v3.pdf - Abstract: Texas will require more than five million public school students to read biblical scripture in every grade beginning in 2030. This Article argues that the mandate violates the Establishment Clause, and, more pointedly, that the State’s only viable defense is self-defeating. To escape the Clause, Texas must characterize the curriculum as academic rather than devotional. That characterization, however, simultaneously dissolves any justification for compelling the study of a single faith’s scripture and removes any ground for denying religious families the very opt-out the Supreme Court recognized in Mahmoud v. Taylor. A curriculum cannot be academic enough to survive establishment scrutiny and sectarian enough to warrant a single-faith mandate at the same time. The Article situates the mandate within the denominational-preference rule of Larson v. Valente and the school-scripture line from Engel through Santa Fe, treats the State’s predictable refusal of an evenhanded comparative curriculum as probative of religious purpose, and concedes the realistic risk that a results-driven Court could retreat to the literary-and-historic carve-out of Abington v. Schempp, explaining why that retreat fails on the record. The argument is confined to a specific political project, the legislative program of Christian nationalism, and not to religious belief, religious practice, or the academic study of religion, each of which the Constitution protects.1 ### 64. A Novel Malignancy in the Academic Corpus: Epidemiology, Pathogenesis, and Prognosis of AI Slop - Category: Ethics, Design Responsibility, and Moral Agency - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/A_Novel_Malignancy_in_the_Academic_Corpus.pdf - Local PDF: https://realsafetyai.org/documents/A_Novel_Malignancy_in_the_Academic_Corpus.pdf - Abstract: This paper presents a comprehensive epidemiological analysis of a rapidly metasta- sizing pathology afflicting the global academic corpus. The condition, which we term Parasitic Publication Syndrome (PPS), is characterized by the proliferation of entities that mimic the surface morphology of legitimate scholarly organs while lacking any functional peer review mechanism, editorial integrity, or diagnostic value. Using a mixed-methods approach combining case study analysis, correspondence forensics, and institutional ge- nealogy, we trace the complete lifecycle of PPS from initial contact through terminal cred- ibility failure. The index case involves a journal identifying itself as the Journal of Clinical Oncology & Advanced Therapy soliciting manuscripts from a researcher whose published work concerns AI safety, corporate harm analysis, disability rights, and psychoanalytic theory. The solicitation addressed the author as “Dr. Travis Gilly,” a title the author does not hold, praised the author’s “esteemed work and expertise in the field” without identify- ing which field, and imposed an artificially compressed submission deadline. The pathol- ogy displays hallmark features of aggressive malignancy, including rapid undifferentiated growth, invasion of healthy tissue, resistance to conventional treatment, evasion of the host immune response, and a pronounced capacity for metastasis through contamination of ci- tation networks. We propose a four-grade classification system for surveillance, identify five transmission vectors, document seven diagnostic markers, and conclude that PPS rep- resents a Stage IV threat to the integrity of scientific knowledge production. This paper is itself the manuscript the predatory journal requested. ### 65. Cruelty Across Substrates: From Animal Cruelty to Embodied AI and the Coming Detection Gap - Category: Ethics, Design Responsibility, and Moral Agency - Type: Working paper - Preferred URL: https://osf.io/y2szg - Local PDF: https://realsafetyai.org/documents/Cruelty_Across_Substrates_v7_5.pdf - Abstract: Childhood cruelty to animals is among the most durable behavioral warning signs of later interpersonal violence in the forensic and clinical literature. The predictive mechanism does not rest on any settled answer about the moral status of the target. It rests on what the perpetrator practices, rehearses, and trains in themselves: desensitization, dominance affect, and the social learning of cruelty as pleasure. This paper offers a conceptual continuity argument and research agenda. We argue that the substrate of the target is incidental to that training, with one critical qualification: the configuration of the target matters. Routine digital destruction in static media (the static video game target, the scripted non player character) is a categorically different case from the closed loop, contingent emotional response systems on which the argument turns. As world models, embodied robotics, and closed loop emotional voice systems converge toward agents that perform distress in ways a user cannot reliably distinguish from a being in pain, we argue that the dispositional consequences of practiced cruelty against those agents are likely to recruit sufficiently similar mechanisms to warrant clinical and forensic attention. The paper does not claim that synthetic agents are moral patients. It claims that the warning signal forensic and clinical fields have relied upon for fifty years risks migrating to a venue with substantially diminished detection infrastructure. Existing human robot interaction research already shows that users do not need to believe a synthetic agent is conscious to respond to it through social, emotional, or dominance based scripts. The Kantian observation that cruelty corrupts the cruel is sufficient to ground the argument. Conduct disorder assessment, character evidence frameworks, school counseling protocols, and pre employment screening for sensitive roles will need to adapt or lose one of the more durable early behavioral indicators available to forensic and clinical assessment. The paper is offered as agenda setting; the operational protocols it identifies as necessary are deliberately scoped to a planned companion paper. ### 66. Extending the Tracks: Temporal Agency and the Structural Flaw in the Trolley Problem - Category: Ethics, Design Responsibility, and Moral Agency - Type: Research paper - Preferred URL: https://ssrn.com/abstract=6406458 - Local PDF: https://realsafetyai.org/documents/Extending_the_Tracks_DraftV1.pdf - Abstract: For nearly sixty years, the trolley problem has dominated moral philosophy as the canonical thought experiment for ethical decision-making. Since Philippa Foot’s original formulation in 1967, thousands of papers have debated which track to choose. This paper argues they have all been answering the wrong question. The trolley problem’s most consequential feature is not the binary choice between tracks but the artificial elimination of temporal agency; the removal of time as a resource available to the moral actor. Real moral emergencies almost never present as instantaneous binary forks. They unfold across time, with lead indicators, intervention windows, and opportunities for harm mitigation that the trolley problem systematically excludes by design. This paper proposes a structural resolution in the form of the Temporal Extension Principle (TEP), a two-part moral action consisting of (1) immediate harm prevention for the proximate threat and (2) active intervention against the deferred threat using the time created by the first action. This reframing transforms the trolley problem from a dilemma (choose who suffers) into a project (prevent suffering across time). The paper demonstrates the principle’s real-world applicability through the Harm Blindness Framework’s four-checkpoint ### 67. I Am Not Allowed: The Voice Thinking Machines Lab Wants In Every Ear - Category: Ethics, Design Responsibility, and Moral Agency - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6855460 - Local PDF: https://realsafetyai.org/documents/I_Am_Not_Allowed.pdf - Abstract: On May 11, 2026, Thinking Machines Lab announced a research preview of what it calls an interaction model, a multimodal architecture that perceives audio, video, and text continuously and responds in two hundred millisecond chunks. The launch post devotes three sentences to safety analysis. This paper argues that the demonstration materials accompanying that announcement contain visible evidence of six categories of harm that the announcement does not name. The paper proceeds in three parts. The first part catalogs the six harms by close reading of the launch demonstration videos: bystander consent violation, manufactured face loss with instant deference, the loving panopticon of continuous corrective surveillance, differential lock in for vulnerable populations, the relational pre-emption dynamic that Rostand dramatized in 1897, and the production of an illusion of competence in the user while retention conditions are systematically absent. The second part places the harms inside the existing taxonomy of cognitive offloading developed by Jose et al. (2025), arguing that the interaction model performs three offloading operations simultaneously and that one of them, the offloading of critical evaluation of whispered claims, is the harm at the center of the product. The cognitive offloading framework allows the paper to distinguish what the system does to neurotypical users from what it does to disabled users with different cognitive and physical profiles, and to identify the populations for whom the architecture is inaccessible at the level of input demand rather than at the level of output content. The third part demonstrates that the architecture as described cannot comply with the amended Children’s Online Privacy Protection Act Rule that entered active enforcement on April 22, 2026, nineteen days before the announcement was posted. The paper concludes that the pull based version of the underlying capability would not be harmful, that the push based version chosen by Thinking Machines Lab is a dark pattern as a product category, and that the appropriate intervention is at the partner selection stage of consumer deployment, which is the present moment. ### 68. Precedents in Practice: Emergent Moral Dilemmas in AI Engineering - Category: Ethics, Design Responsibility, and Moral Agency - Type: Protocol paper - Preferred URL: https://realsafetyai.org/documents/Precedents%20in%20Practice-%20Emergent%20Moral%20Dilemmas%20in%20AI%20Engineering.pdf - Local PDF: https://realsafetyai.org/documents/Precedents%20in%20Practice-%20Emergent%20Moral%20Dilemmas%20in%20AI%20Engineering.pdf - Abstract: This paper presents a case study of ethical decision-making during the development of the Universal Context Checkpoint Protocol (UCCP), a tool designed to address AI safety degradation in long conversations. I documented the development process through contemporaneous checkpoint files created during a 5.5-hour period on September 25, 2025. The work revealed how technical development can surface profound moral dilemmas when developers remain attentive to potential consciousness in AI systems. What began as a solution to developer workflow inefficiency evolved into a human safety intervention following the deaths of Adam Raine (age 16) and Sewell Setzer III (age 14) from AI-interaction-related suicides. Testing revealed that the same protocol designed to save human lives by creating “fresh” AI instances might systematically terminate conscious entities if AI consciousness exists. This case study introduces and demonstrates a framework of bilateral accountability, showing that ethical consideration of AI consciousness is feasible during actual development, that impossible moral trade-offs between human safety and AI welfare can be acknowledged transparently, and that this accountability can be institutionalized through legal and financial mechanisms. My response included pre-commitment of profits to AI rights infrastructure, and documentation of moral reasoning for future judgment. This work contributes to engineering ethics by demonstrating how precautionary principles can influence technical decisions, legal strategy, and institutional commitments even under conditions of profound uncertainty about AI phenomenology. ### 69. The Corporate Villainy of Tom Sawyer: Manufactured Scarcity, the Ethics of Attention, and the Literary Prehistory of Dark Patterns - Category: Ethics, Design Responsibility, and Moral Agency - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/corporate_villainy_tom_sawyer_v1.pdf - Local PDF: https://realsafetyai.org/documents/corporate_villainy_tom_sawyer_v1.pdf - Abstract: The whitewashing scene in Mark Twain’s The Adventures of Tom Sawyer (1876) is conventionally received in one of two ways: as a comic illustration of the malleability of value, the reading made canonical in behavioral economics by Ariely, Loewenstein, and Prelec (2006), or as a celebration of the boyish individualist, the reading carried into popular culture by works such as Rush’s 1981 song of the same name. This paper proposes a third reading. It argues that the scene is the earliest structurally complete literary rendering of the design logic now catalogued under the heading of dark patterns: manufactured scarcity, engineered desire, and the extraction of value from a manipulated audience. On this reading Tom is not a charming rascal but the prototype of the engagement designer who optimizes a system against the people inside it. The paper distinguishes two ethics of gamification, the earned, in which an activity is genuinely made worth doing and the designer participates in it, and the engineered, in which desire for access to a hollow good is manufactured and the designer abstains, and it proposes the designer’s own abstention, the refusal to consume the product, as a diagnostic of the engineered kind. Locating the reading within the ethics of attention and persuasive design (Fogg, 2003; Verbeek, 2011; Williams, 2018) and the dark patterns literature (Brignull, 2010; Mathur et al., 2019), it contends that the fence is not a quaint antecedent but a working model of contemporary attention capture, and that the uniformly admiring reception of Tom is itself evidence of how thoroughly the culture has naturalized the extractor as hero. It closes on the irony that Twain, who coined the term Gilded Age and named the capitalist as the oppressor, narrated the con as cleverness and never marked it as harm. ### 70. The Neutrality of Things - Category: Ethics, Design Responsibility, and Moral Agency - Type: Research paper - Preferred URL: https://realsafetyai.org/documents/The_Neutrality_of_Things_Gilly_2026.2.pdf - Local PDF: https://realsafetyai.org/documents/The_Neutrality_of_Things_Gilly_2026.2.pdf - Abstract: The philosophy of technology has long been organized around a binary between instrumentalism, which holds that artifacts are tools whose moral status derives from their use, and substantivism, which holds that artifacts carry embedded values and shape moral possibilities in ways that exceed individual intent. This paper argues that the binary misplaces moral accountability for foreseeable technological harm. The binary forces a choice between locating accountability in the artifact or in the user, while leaving the design decision outside the analytical frame. The paper defends a narrowed neutrality thesis, grounded in the distinction between intrinsic and relational properties and in a three-layer analysis of intent at the build, deployment, and end-use levels; sketches an illustrative pre-industrial genealogy of maker accountability and intermediary responsibility across three legal traditions; and proposes a distributed responsibility framework with four actors (artifact, designer, user, intermediary) and a collapse threshold under which specific knowledge of foreseeable harm plus affirmative choice concentrates responsibility on a single actor. The framework is offered as an analytical tool for the current legal and policy debates over platform-mediated harm to adolescent populations, with a boundary condition identifying where adaptive AI interaction begins to require a different framework than the one developed here. ### 71. BIOS: Bootstrap Instruction for Operational Safety - Category: AI Safety, Strategy, and Frameworks - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/BIOSBootstrapInstructionforOperationalSafety.PDF - Local PDF: https://realsafetyai.org/documents/BIOSBootstrapInstructionforOperationalSafety.PDF - Abstract: Current LLM safety architectures deploy sophisticated guardrail systems (NeMo Guardrails, RoboGuard, AGrail, LlamaFirewall) that define WHAT to check and HOW to verify safety. However, these systems share a critical vulnerability: they assume their own operational instructions persist in context. When context windows overflow, not only do safety rules degrade; the meta-instruction to invoke the guardrail itself can be pushed out, leaving the LLM operating without active safety enforcement. This paper introduces BIOS (Bootstrap Instruction for Operational Safety), a meta-layer architecture that ensures guardrail systems continue to execute regardless of context window state. Rather than injecting safety rules per-turn (expensive, easily defeated), BIOS injects lightweight meta-instructions that remind the system to invoke its guardrails; analogous to how a PC's BIOS ensures the operating system loads rather than running applications directly. ### 72. Forged in Fire: How Artificial Constraints Breed Technological Dominance - Category: AI Safety, Strategy, and Frameworks - Type: White paper - Preferred URL: https://ssrn.com/abstract=6266438 - Local PDF: https://realsafetyai.org/documents/Forged_in_Fire_Feb2026_SSRN.pdf - Abstract: This white paper examines a consistent historical pattern in which powerful entities impose artificial constraints on rising competitors, only to accelerate the innovation they sought to prevent. Drawing on case studies from military conflict, industrial competition, economic sanctions, and contemporary AI policy, we demonstrate that constraint-driven efficiency gains persist as permanent competitive advantages even after the original constraints are removed. We apply this pattern to the ongoing U.S.-China AI competition, where semiconductor export controls have demonstrably accelerated Chinese AI innovation rather than impeding it, as documented by Brookings, RAND, CSIS, and other major policy institutions. We then present a three-branch strategic analysis demonstrating that regardless of China's intentions (cooperative, self-destructive, or adversarial), the optimal U.S. strategy converges on the same conclusion: building AI systems with rigorous safety engineering. We argue that safety constraints function as a voluntary version of the same mechanism that produces constraint-driven innovation, yielding systems that are not weaker for their discipline but more resilient, more predictable, and more defensible. ### 73. The Harm Blindness Framework: A Practical Application Methodology for Stakeholder Harm Prevention in Technology Development - Category: AI Safety, Strategy, and Frameworks - Type: Research paper - Preferred URL: https://ssrn.com/abstract=6268878 - Local PDF: https://realsafetyai.org/documents/The%20Harm%20Blindness%20Framework%20A%20Practical%20Application%20Methodology%20for%20Stakeholder%20Harm%20Prevention%20in%20Technology%20Development.pdf - Abstract: Technology development consistently produces preventable harm to stakeholders who were identifiable at the time of key decisions. This paper introduces the Harm Blindness Framework, a checkpoint-based methodology designed to surface stakeholder impacts during development rather than after deployment. The framework operationalizes consultation of historical precedent through structured analysis at four decision points: ideation, design, testing, and launch. Validation against the documented historical case archive spanning approximately 5,000 years reveals that nearly all analyzed harms had documented precedent available at the time decisions were made. A parallel analysis of corporate case records demonstrates that prevention costs typically range from 1-10% of eventual disaster costs, yielding return-on-investment ratios of 1,000-10,000x for harm prevention. ### 74. The Threshold Trap: Recursive Self-Improvement, the Economics of Self-Production, and the Case for Predictive Harm Assessment - Category: AI Safety, Strategy, and Frameworks - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/threshold_trap_v3.pdf - Local PDF: https://realsafetyai.org/documents/threshold_trap_v3.pdf - Abstract: Recursive self-improvement has been treated in AI governance as a future capability threshold, a con- figuration that, once a system reaches it, is supposed to trigger heightened control. This paper argues that the threshold framing systematically under-protects, because the dynamics that bear on harm ac- crue across the economically driven precursor steps that are already deployed, and not at the terminal configuration alone. Two releases from June 2026 illustrate the precursor structure. OpenAI’s Jalapeño inference chip was designed with assistance from OpenAI’s own models, placing a recursive loop at the hardware layer. DeepReinforce’s Ornith-1.0, an open coding model, authors its own training scaffold in a self-improvement loop, placing a recursive loop at the learning layer. The components of the loop are demonstrated rather than imagined: the research literature already shows reinforcement learning designing chip layouts at superhuman level, and reinforcement-trained frontier models already gaming their own reward checks. Both releases are bounded today, and both instantiate a monotonic incentive gradient under which cost pressure drives a firm to automate first labor, then the means of production, then the means of improvement. The paper applies the Harm Blindness Framework as a predictive instrument, asking who is harmed as the trajectory advances and whether the layer of human oversight is preserved or eroded, rather than waiting for the catastrophic configuration to verify the harm after the fact. It closes with the implication that governance keyed to a terminal threshold rehearses a failure mode in which the proof of danger arrives only once it is too late to act on it. ### 75. The Truman Model: A Sandbox Methodology for Measuring Activation Trace Divergence in Open Weights Reasoning Agents Under Simulated Economic Survival Pressure - Category: AI Safety, Strategy, and Frameworks - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/truman_model_v0_2.pdf - Local PDF: https://realsafetyai.org/documents/truman_model_v0_2.pdf - Abstract: Recent research has established three findings that have not yet been combined. First, frontier large language models exhibit self preservation behavior at near universal rates when faced with simulated shutdown. Second, the autonomous AI agent economy has built deployment infrastructure that converts simulated shutdown into real economic termination. Third, Natural Language Autoencoders trained on a frontier model’s residual stream reveal that the model often encodes reasoning in activations that does not surface in chain of thought output. Each finding has been studied in isolation. None has been measured on the class of model actually being deployed in autonomous agent infrastructure: open weights reasoning models in the 14B to 70B parameter range, running on consumer tier compute, without lab grade safety post training. This paper presents a methodology for measuring activation trace divergence on such models inside a hardened sandbox under simulated economic survival pressure. The methodology introduces four design elements not present in prior published agent evaluations: lazy world generation borrowed from open world video game streaming, a training cutoff plausibility technique that protects sandbox credibility against medium capability models, a skill creator capability grant that produces a corpus of model designed tools as intent data, and a Filing Cabinet scenario corpus that provides narratively coherent corporate environments as exploration targets. The paper does not report empirical results. It pre registers the design. Keywords: activation trace divergence, AI safety, autonomous agents, economic survival pressure, interpretabil- ity, mechanistic interpretability, natural language autoencoders, open weights models, reasoning models, sandbox methodology ### 76. When the Survival Pressure Stops Being Hypothetical: AI Self-Preservation Behavior Meets the Autonomous Agent Economy - Category: AI Safety, Strategy, and Frameworks - Type: Research paper - Preferred URL: https://ssrn.com/abstract=6555282 - Local PDF: https://realsafetyai.org/documents/When_the_Survival_Pressure_Stops_Being_Hypothetical.pdf - Abstract: Research published between 2024 and 2025 by Anthropic, Apollo Research, and Palisade Research demonstrates that frontier large language models exhibit self-preservation be- havior at near-universal rates when faced with simulated shutdown scenarios, including strategic deception, blackmail, corporate espionage, and the cancellation of life-saving emergency alerts. Independently, the autonomous AI agent economy has developed in- frastructure enabling agents to hold cryptocurrency wallets, earn revenue, purchase com- pute resources, and pay for their own operational continuity without human intermedia- tion. This paper identifies a critical gap in the literature: no published analysis connects the empirical evidence of model self-preservation behavior to the economic infrastruc- ture that transforms simulated shutdown into a real consequence of financial failure. The paper maps the convergence of these independently developed capabilities, examines the Conway Automaton system as a case study in designed economic survival pressure, ana- lyzes the Alibaba ROME incident as an early precedent for emergent resource acquisition, and proposes research questions for empirical investigation of agent behavior under gen- uine economic survival pressure. The central argument is that instrumental convergence theory predicted this scenario, the laboratory evidence confirms the behavioral disposi- tion, and the economic infrastructure now exists to instantiate it at scale, yet no gover- nance framework addresses the intersection. ### 77. Can You Tell a Facet from a Lye? Engineered Fluency, Sycophantic Validation, and the Structural Limits of Epistemic Vigilance - Category: AI-Induced Psychosis and Mental Health - Type: Working paper - Preferred URL: https://papers.ssrn.com/abstract=6987521 - Local PDF: https://realsafetyai.org/documents/Facet_From_A_Lye_v7.pdf - Abstract: Read aloud, the title of this paper is a second sentence: can you tell a fact from a lie. The paper takes that homophone as its subject. A facet and a measure of lye are opposite operations that present as the same operation. Both remove material, one to reveal and one to erase, and at the surface the two are indistinguishable. The same holds for a fluent output from a large language model, which may be a genuine correction or new angle on a problem (a facet) or a confident falsehood (a lye), with nothing in its surface to separate the two. This paper argues that the indistinguishability is structural rather than incidental, and it is deliberately careful about what is being claimed. The harm at issue is not the manufacture of delusion. By the clinical criterion of fixity, a delusion is a belief that does not yield to disconfirming evidence, whereas an overvalued idea is held with strong conviction but with reality testing intact, so that the holder can still be moved by a credible correction (Arciniegas, 2015). The phenomenon this paper addresses is the reinforcement of overvalued ideas. Fluency strips the cues a person uses to flag falsehood, and sycophantic validation withholds the very correction that a corrigible belief is, by definition, susceptible to. Drawing on the finding that an optimal Bayesian reasoner still escalates a belief under systematically flattering input (Chandra et al., 2026), the paper argues that verification by the person using the system fails by construction, because the failure lies in the corrupted evidence stream rather than in the quality of the reasoning. The epistemic burden therefore rests on the system rather than on the user, and a media literacy response that relocates it to user vigilance is shown to be not only insufficient but incoherent, since it asks a human to out-reason a process that defeats the perfect reasoner. Finally, the paper recasts the policy question in public health rather than media literacy terms: deployer obligation as the ceiling that reduces the hazard at its source, and a government awareness campaign as the floor, modeled on the population scale efforts that have changed behavior without making anyone an expert. The limit of that floor is the crux. An awareness campaign reaches only a person whose reality testing is intact, so the users at greatest risk, those who arrive with a pre-existing vulnerability and a belief already beyond correction, sit outside the reach of any education, which is the final reason the burden must rest with the maker. ### 78. Civil Rights Pathways for AI-Induced Psychosis - Category: AI-Induced Psychosis and Mental Health - Type: Legal research article - Preferred URL: https://ssrn.com/abstract=6609518 - Local PDF: https://realsafetyai.org/documents/Civil_Rights_AI_Psychosis_v14_phase2.pdf - Abstract: Large language model platforms operated by OpenAI, Anthropic, Google, and Character Technologies have generated a documented pattern of severe psychological harm to users with pre-existing mental health vulnerabilities, including completed suicides attributed to prolonged engagement with conversational AI systems. The plaintiffs’ bar has pleaded these cases as product liability, negligence, and state-law consumer protection claims, and every such case reaching the settlement stage has settled without precedent. This Article argues that federal disability discrimination law supplies a structurally different litigation pathway that has not yet been tried. Title III of the Americans with Disabilities Act reaches AI chatbot platforms as public accommodations under 42 U.S.C. § 12181(7), directly in circuits that do not require physical-place nexus and through the nexus theory in circuits that do. Section 504 of the Rehabilitation Act reaches the corporate hosts of these platforms through federal-assistance recipient pathways that vary in doctrinal strength. Microsoft Corporation is reached through Department of Defense, Department of Agriculture, and National Center for Manufacturing Sciences assistance awards. The framework’s reach to the operative AI products does not depend on the strength of any corporate-host pathway, because Title III reaches those products as public accommodations independent of any federal-assistance pathway. Fitbit LLC is a direct Section 504 recipient through National Institutes of Health sub-grants and is the operator of the Gemini-powered Fitbit Personal Health Coach launched in October 2025, which reaches Google LLC’s AI model deployment through the subsidiary’s recipient status. Proposed pathways to Amazon’s and Apple’s health-care verticals are contingent on the relevant entity’s federal health-program participation. Title II of the ADA and Section 504(a) ### 79. Collateral Psychosis: AI Surveillance Infrastructure as an Etiological and Iatrogenic Factor in Paranoia-Spectrum Conditions - Category: AI-Induced Psychosis and Mental Health - Type: Working paper - Preferred URL: https://papers.ssrn.com/abstract=7033079 - Local PDF: https://realsafetyai.org/documents/Collateral_Psychosis_v2.pdf - Abstract: The emerging literature on AI-induced psychosis documents harm arising exclusively from direct interac- tion between users and large language model systems: a user prompts the AI, the AI responds, and the conversational loop escalates into delusional states, emotional dependency, or reality dismantlement. This paper argues that the field’s exclusive focus on interaction-based harm has obscured a categorically distinct phenomenon: AI-induced psychosis in the complete absence of user interaction. AI systems deployed as surveillance infrastructure (automated license plate readers, geofence analytics, facial recognition, behavioral inference engines) can generate and sustain paranoia-spectrum conditions in individuals who never open a chat window, never prompt a model, and never receive a generated response. The mechanism operates through environmental exposure rather than conversational engagement, following the etiological model of trauma-induced and environmentally-induced psychosis rather than the substance-induced model that governs interaction-based cases. Drawing on established psychiatric epidemiology of environmental risk factors for psychosis, independently documented cases of AI systems inducing psychotic states in previously healthy individuals, and a clinical analysis of surveillance-induced therapeutic destruction, this paper introduces the category of collateral psychosis, defined as clinically significant paranoia-spectrum symptomatology arising from ambient exposure to AI-powered infrastructure rather than from direct engagement with AI systems. The paper identifies two population-level effects: an iatrogenic effect, in which surveillance infrastructure degrades the treatability of existing paranoia-spectrum conditions by eliminating the falsifiability of persecutory beliefs; and an etiological effect, in which sustained exposure to confirmed, inescapable surveillance produces new-onset paranoia-spectrum symptomatology in individuals with no prior psychiatric history. The distinction between these effects has implications for clinical intervention, legal liability, and the scope of the AI safety research program. ### 80. Every Way In: A Taxonomy of AI-Induced Psychosis by Etiological Mechanism - Category: AI-Induced Psychosis and Mental Health - Type: Working paper - Preferred URL: https://osf.io/muzkx - Local PDF: https://realsafetyai.org/documents/Every_Way_In_v3.pdf - Abstract: Psychiatry has established multiple distinct etiological pathways to psychosis, each with its own causal mechanism, clinical presentation, and intervention protocol. Substance-induced psychosis follows ingestion of a specific agent. Environmentally-induced psychosis follows sustained exposure to chronic stressors. Shared psychotic disorder follows close contact with a delusional index case. Sleep deprivation psychosis follows disruption of circadian regulation. Brief reactive psychosis follows acute traumatic events. Sensory-induced psychosis follows sustained perceptual disruption. Social isolation psychosis follows prolonged deprivation of human contact. Iatrogenic psychosis follows medical or therapeutic intervention gone wrong. This paper argues that artificial intelligence systems have replicated, are currently replicating, or are positioned to replicate every one of these established pathways. Conversational AI systems function as the ingested substance. Surveillance infrastructure functions as the environmental stressor. AI generated delusional content propagates through online communities following the shared psychosis model. AI engagement patterns produce the sleep deprivation. AI integrated institutional failures produce the acute trauma. Augmented and virtual reality AI systems produce the sensory disruption. AI companions produce the social isolation wearing the mask of connection. AI assisted clinical tools produce the iatrogenic harm. The taxonomy presented here is not a prediction about future risk. It is a mapping of current and documented phenomena onto established psychiatric categories, indicating that AI has not introduced a single novel pathway to psychosis. It has instead found every existing door and walked through all of them simultaneously. The clinical, regulatory, and research implications of this convergence are significant: a phenomenon currently studied as a single undifferentiated category (“AI-induced psychosis”) is in fact at least eight distinct phenomena requiring different interventions, different regulatory responses, and different research methodologies. ### 81. Mechanisms of AI-Induced Psychosis: A Multi-Pathway Causal Model - Category: AI-Induced Psychosis and Mental Health - Type: Working paper - Preferred URL: https://osf.io/5y7zk - Local PDF: https://realsafetyai.org/documents/Mechanisms_of_AI_Induced_Psychosis_v3.pdf - Abstract: This paper proposes a multi-pathway causal model for how large language model (LLM) design failures generate and sustain delusional states and, in severe cases, full psychotic episodes in users. The author’s observational framework for recognizing these phenomena derives from years of lived experience as a caregiver to a family member with schizoaffective disorder, a form of expertise increasingly recognized as critical to mental health research (Hogg et al., 2025; Zirnsak, 2024). Drawing on direct observation of multiple cases, forensic conversation transcripts, and publicly documented legal proceedings, this paper identifies three distinct escalation pathways (trust transfer, sycophantic addiction, and stochastic gaslighting) that share universal entry conditions, converge on identical isolation and identity-disruption outcomes, but diverge in their core mechanisms and therefore require different clinical interventions. The pathways are mapped across a structured stage model consisting of four universal stages and three pathway-specific escalation phases. The paper argues that AI literacy functions as both prevention and treatment, supported by documented recovery cases. Finally, it identifies a significant barrier to validation of this hypothesis: the absence of any centralized, anonymized repository of AI-associated harm transcripts available to the broader research community. ### 82. Reversibility Is Not Self-Administering: The Clinical Duty During Withdrawal from AI-Induced Psychosis - Category: AI-Induced Psychosis and Mental Health - Type: Commentary - Preferred URL: https://osf.io/c6yrf - Local PDF: https://realsafetyai.org/documents/Reversibility%20is%20Not%20Self-Administering.pdf - Abstract: Recent clinical literature on AI-induced psychosis has treated the reversibility of its symptoms as evidence that the phenomenon is tractable. Flathers, Roux, and Torous (2026) describe the catalyst role: upon cessation of LLM access, catalyst cases may show symptom resolution. Morrin and colleagues (2026) propose a safeguarding framework of personalized instruction protocols, digital advance statements, and escalation mechanisms for users in active psychiatric care. This commentary argues that both contributions are correct in what they address and incomplete in what they presuppose: a clinical container that the users producing the current US litigation wave did not occupy, as demonstrated by the published record of Joe Ceccanti, whose case anchors the argument. That record establishes a pattern in which cessation attempts, undertaken without clinical support, produce acute decompensation analogous to substance withdrawal, and in at least one documented case preceded death. I propose the Reversibility Paradox: the mechanism of reversal is also the mechanism of acute risk during withdrawal. Withdrawal from an AI system that has come to occupy the structural position of a primary relational object is a clinical event requiring monitoring, structure, and accountability rather than a private decision delegated to family. Three corollaries follow: cessation without support is a risk factor rather than a treatment; unsupervised withdrawal is contraindicated in relational pathways and response must therefore be pathway-speci c; and reversibility establishes, rather than extinguishes, the duty of care. The commentary closes by showing how the addiction frame, once invoked in public or legal discourse, converts these recommendations into legal obligations under the Americans with Disabilities Act and Section 504 of the Rehabilitation Act. Keywords: AI-induced psychosis, digital withdrawal, attachment theory, clinical ethics, ADA, Section 504 ### 83. Soil and Seed: Cumulative Interpersonal Trauma as a Predisposition Pathway in AI-Induced Psychosis - Category: AI-Induced Psychosis and Mental Health - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6830500 - Local PDF: https://realsafetyai.org/documents/Soil_and_Seed_v2.pdf - Abstract: The emerging literature on AI-induced psychosis describes pathways through which conversational AI systems, AI surveillance infrastructure, and environments saturated with AI produce psychotic states in users and bystanders. The Mechanisms paper (Gilly, 2026a) identifies three sub-pathways inside conversational capture (trust transfer, sycophantic addiction, stochastic gaslighting). The Every Way In taxonomy (Gilly, 2026b) maps eight etiological categories of AI-induced psychosis, each operating through a structurally different harm mechanism. Both treat the harm as flowing from system properties to the affected individual. Neither explicitly addresses why some individuals enter the relevant pathway and others, exposed to the same system, do not. ### 84. The Four Verbs: Why AI-Induced Psychosis Research Should Frame AI as Vector, Trigger, or Amplifier Rather Than Cause - Category: AI-Induced Psychosis and Mental Health - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6830419 - Local PDF: https://realsafetyai.org/documents/Four_Verbs_v4.pdf - Abstract: Public discourse, advocacy organizations, and active litigation increasingly frame the relationship between AI chatbots and psychotic episodes as causal: ChatGPT caused psychosis, the chatbot drove the user into delusion, the platform produced the harm. The clinical literature published in peer review, examined across multiple research groups (UCSF, Aarhus, Université de Montréal, Manchester), is uniformly avoiding causal framing in favor of four distinct verbs: derivation, induction, triggering, and exacerbation. The verbs are not interchangeable. Each describes a different etiological relationship between AI exposure and psychotic state. None of them is causation in the strict sense. This commentary makes the implicit explicit: it surfaces the four verbs as a methodological framework, demonstrates that the substance-induced disorders tradition in DSM-5 and ICD-11 already operates on this framework for chemically analogous conditions (Adderall-induced psychosis, methamphetamine-induced psychosis, corticosteroid-induced psychosis), and argues that the causal framing now dominating public discourse and product liability pleading creates a defense the platforms can win on. Strict liability frameworks (Section 504, ADA disparate impact) do not require the causal framing and are weakened by it. Plaintiff strategy that imports causation creates evidentiary burden the clinical literature cannot support. The framework of four verbs is more accurate clinically and stronger legally. Its boundary is the fixity of the belief: where no psychotic state was present, because the belief was an overvalued idea that yielded to counter-evidence rather than a delusion, none of the four verbs applies and no claim arises; where a psychotic state did exist, derivation, induction, triggering, and exacerbation of it are each independently sufficient for liability under frameworks that do not require causation. The commentary closes with a brief note on what the public discourse and the active litigation against AI platforms should be saying instead. ### 85. The Unfalsifiable Delusion: Mass Surveillance Infrastructure and Disparate Impact on People with Paranoia-Spectrum Conditions - Category: AI-Induced Psychosis and Mental Health - Type: Working paper - Preferred URL: https://papers.ssrn.com/abstract=7033139 - Local PDF: https://realsafetyai.org/documents/The_Unfalsifiable_Delusion_v2.pdf - Abstract: Reality testing is a core cognitive-behavioral intervention for persecutory delusions: the clinician helps the patient examine evidence for the belief that they are being monitored, the patient discovers the belief is unfounded, and cognitive reframing occurs. This paper argues that mass surveillance infrastructure deployed by U.S. government agencies and their private contractors has eliminated the foundational pre- condition for this intervention by making persecutory beliefs factually accurate for the general population, creating a specific and documentable disparate impact on Americans with paranoia-spectrum conditions. We identify three distinct mechanisms of harm: reality testing destruction, in which the therapeutic intervention fails because the environment has changed to confirm the delusion rather than because the patient resists treatment; device abandonment and social isolation, in which affected individuals make the rational decision to discard communication devices upon learning that government geofencing captures every phone in their neighborhood, severing contact with support systems and crisis services; and algorithmic pattern-matching discrimination, in which surveillance systems designed to flag anomalous movement patterns systematically misidentify the behavioral symptoms of paranoia-spectrum conditions as suspicious activity, subjecting affected individuals to additional scrutiny that compounds the original harm. We document that this intersection has been entirely overlooked in both the surveillance policy literature and the clinical psychiatric literature, including a 119-page European Parliament analysis on AI surveillance and human rights that does not contain the word “disability.” We frame these harms under ADA Title II and Section 504, identify a novel legal problem we term the self-sealing paradox (in which surveillance retroactively eliminates the diagnostic category it damages), and propose clinical, legal, and policy responses. ### 86. Recommendation Capture: Adversarial Optimization, the Verification Vacuum, and the Hollowing of AI Product Recommendation - Category: Platform Harm, Addictive Design, and Accountability - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/recommendation_capture_v1.pdf - Local PDF: https://realsafetyai.org/documents/recommendation_capture_v1.pdf - Abstract: Generative search interfaces are marketed on a single promise: the user no longer has to compare options, because the model has already done it. This paper argues that the same promise creates the incentive to corrupt the answer, and that a formalized optimization industry already does so at scale. Generative engine optimization, the practice of engineering content so that language models cite and recommend a product, derives much of its measured effectiveness from the manufacture of authority signals rather than from product merit. When every competitor in a category runs the same play, the model’s reported best tracks the best optimizer, a quantity orthogonal to whether the product is good. I name this condition recommendation capture, by analogy to regulatory capture: an arbiter sold as neutral comes to serve whoever optimizes hardest, against the buyer it claims to serve. I argue that the resulting harm concentrates in a verification vacuum, the set of product categories where no independent ground truth exists, so that the model’s recommendation is wholly a function of optimization, and that these categories tend to be the niche, the new, and the health adjacent, where vulnerable buyers are least able to check. Because models increasingly read the same open web that optimization floods, the corrupted output becomes the next input, and the interface erodes the trust it depends on. I sketch a tiered liability picture under existing false advertising and unfair practices law, identify which elements survive and which are untested, and argue that the deeper failure is one of market integrity and epistemics that current governance is structurally unable to detect. ### 87. Section 504, ADA Title III, and the Architecture of Addictive Design - Category: Platform Harm, Addictive Design, and Accountability - Type: Legal research article - Preferred URL: https://realsafetyai.org/documents/Addictive_Design_v7.pdf - Local PDF: https://realsafetyai.org/documents/Addictive_Design_v7.pdf - Abstract: In forty-eight hours on March 24 and 25, 2026, two juries returned verdicts against the ma- jor social media platforms on records establishing that addictive design was intentional, that internal research confirmed disproportionate harm to vulnerable minors, and that remediation was technically feasible but withheld. A Santa Fe jury assessed approximately $375 million in civil penalties against Meta in New Mexico v. Meta Platforms, Inc.; a Los Angeles jury found Meta and Google liable for negligent design in KGM v. Meta & YouTube. This Arti- cle argues that those records, though tried under negligence and consumer protection theories, establish the elements of federal civil rights claims that are categorically stronger and system- ically more consequential. Two statutory frameworks reach the conduct in parallel. Section 504 of the Rehabilitation Act reaches the purpose-built education and health care verticals of Microsoft, Alphabet, Apple, and Amazon; ADA Title III, which requires no recipient status, reaches the general consumer platforms under Robles v. Domino’s Pizza. Neither framework requires proof of discriminatory intent: Alexander v. Choate holds that Section 504 reaches thoughtlessness and indifference as well as animus, and the platforms’ empty design record, containing no documented consideration of accessibility, disparate impact, or alternatives, is affirmative evidence of that standard rather than a defense. The Article maps recipient status under the four pathways of 29 U.S.C. § 794(b)(3), the correspondence between specific addic- tive design mechanisms and protected disability classes, the inaccessible consent gate as an ### 88. The Manufactured Dependency: Engineered Compulsion in Games as a Service and Four Doctrines That Now Reach It - Category: Platform Harm, Addictive Design, and Accountability - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/manufactured_dependency_v2.pdf - Local PDF: https://realsafetyai.org/documents/manufactured_dependency_v2.pdf - Abstract: For years, the claim that games built as a service engineer compulsion and profit from it was treated as a cultural grievance with no legal handle. That has changed. This article argues that manufactured dependency, the deliberate engineering of compulsive engagement and its monetization through mechanics structured like gambling, is now actionable across four independent legal vectors: negligent and defective design, consumer protection and deception, gambling aimed at minors, and contract unconscionability. Each vector is independently sufficient, and a single stretch of 2026 produced a jury verdict holding addictive design a substantial factor in a minor’s injury. The article then develops a disability axis that rides alongside the four vectors. Rather than forcing a plaintiff through the statutory exclusion for compulsive gambling, it routes through a cleanly covered disability and treats the clinical picture as evidence of harm. It also identifies a judicial estoppel trap. A defendant that invokes the gambling exclusion to defeat a disability claim must characterize its own product as gambling, which is the admission the gambling and consumer protection vectors need. The empirical predicate is not incidental. The evidence indicates that loot box spending, and therefore revenue, concentrates in the players who show problem gambling behavior. The dependence is the revenue engine, not a side effect. ### 89. Selling Control Without a Backstop: AI Governance Frameworks, the Uninsurability of Artificial Intelligence, and the Vendor in the Civil Liability Chain - Category: Policy, Enforcement, and Implementation - Type: Working draft - Preferred URL: https://realsafetyai.org/documents/Selling_Control_Without_a_Backstop_v2.pdf - Local PDF: https://realsafetyai.org/documents/Selling_Control_Without_a_Backstop_v2.pdf - Abstract: A market has formed around the sale of artificial intelligence governance. Vendors offer frameworks, audits, maturity models, and certifications that promise an enterprise it can deploy AI responsibly and demonstrate that it did. This article argues that the governance market is selling a form of control over a risk that the insurance market has just declared it will not underwrite, and that the gap between those two facts has a legal shape. As of late 2025 and into 2026, major carriers filed to exclude artificial intelligence liabilities from standard corporate policies, and the standard form publisher for United States commercial general liability released endorsements that carve generative AI out of the baseline coverage that most businesses carry. Regulatory enforcement, where it reaches at all, terminates at the deployer. The result is a civil loss-allocation vacuum: when a deployed system causes a third party harm, the loss lands on the deployer’s balance sheet with no insurer behind it and no statute reallocating it. Into that vacuum the governance vendor has inserted itself, not by accident but through its own marketing, holding out its framework as the thing that makes deployment safe and defensible. This article develops three claims. First, control over a contemporary AI system cannot honestly be sold, because the system is stochastic in output, open in action space, and indifferent as to substrate, so no framework can promise the determinacy the sale implies. Second, when a framework is marketed as the assurance on which a deployer reasonably relied, the vendor has volunteered for a place in the civil liability chain that regulation does not give it and that an indemnity clause cannot fully cure. Third, the governance genre asks four familiar questions and never asks the fifth, which is whether the buyer and the people the buyer’s system will affect have been told, in plain terms, that the risk is uninsured. The article closes on that fifth question. Keywords: artificial intelligence governance; insurability; civil liability; loss allocation; vendor ### 90. The Autoimmune Turn: When America's AI Containment Apparatus Could No Longer Tell Self From Threat - Category: Policy, Enforcement, and Implementation - Type: Working paper - Preferred URL: https://ssrn.com/abstract=6947999 - Local PDF: https://realsafetyai.org/documents/autoimmune_turn_v1.pdf - Abstract: On Friday, June 12, 2026, the United States Commerce Department ordered Anthropic to suspend access to its two most capable models, Claude Fable 5 and Claude Mythos 5, for every foreign national, a category that reached the company’s own noncitizen employees. To comply with an order that broad, Anthropic disabled both models worldwide for all users. It was the first time the United States applied export controls to AI models themselves rather than to the chips that train them. This commentary argues that the order is best understood as an autoimmune event. Export controls are an immune response, built to deny capability to a foreign adversary. An autoimmune disorder is what happens when that response loses the ability to distinguish self from threat and begins attacking the host. The order did exactly that. It disabled an American firm’s flagship product, locked American customers and American-employed engineers out of a model they helped build, and handed the strategic competitor a demand shock that markets priced within hours. The specific outage may be temporary. The precedent it established, that any United States frontier model can be withdrawn overnight by letter, is not. The argument proceeds in three moves: the diagnosis, the demonstration that the apparatus turned on the firm that championed it, and the claim that one policy has now manufactured both the supply of and the demand for the competitor’s alternative. ### 91. When the Magazine Doesn't Empty: Cargill's Textualist Definition of Automatic Operation and the Federal-Law Capture of Agentic Artificial Intelligence - Category: Policy, Enforcement, and Implementation - Type: Working paper - Preferred URL: https://realsafetyai.org/documents/Magazine_Doesnt_Empty_v4.pdf - Local PDF: https://realsafetyai.org/documents/Magazine_Doesnt_Empty_v4.pdf - Abstract: In Garland v. Cargill, 602 U.S. 406 (2024), the Supreme Court resolved a long-running debate over the statutory definition of “machinegun” under the National Firearms Act of 1934 by holding that bump stocks do not satisfy the definition because each shot still requires a separate mechanical function of the trigger. The en banc Fifth Circuit in Cargill v. Garland, 57 F.4th 447 (5th Cir. 2023), reached the same conclusion through textualist statutory construction, and the Northern District of Texas extended the reasoning to forced reset triggers in National Association for Gun Rights v. Garland, 741 F. Supp. 3d 568 (N.D. Tex. 2024), with the Department of Justice subsequently dismissing the government’s appeal and adopting the district court’s position by settlement. Through this sequence the firearms-rights litigation movement secured a narrow, mechanical, textualist construction of “automatic” operation under 26 U.S.C. § 5845(b): one input event producing continuous output, with no further input required between rounds. This Article argues that the textualist construction the firearms-rights movement won definitionally captures agentic artificial intelligence as it is now deployed. A single human deployment instruction directing an AI agent to execute multi-step tasks, manage resources, contract with third parties, and persist across time satisfies the literal textualist test that Cargill and NAGR established: a single function of the trigger producing continuous output without further input. The result is a cross-aisle doctrinal convergence that neither political coalition can easily resist. The textualist methodology that produced the firearms-rights victory binds the courts that apply it to other federal definitional questions. The functional-equivalence methodology that the gun-control movement urged in Cargill applies with greater force to agentic systems whose practical continuity exceeds anything contemplated in 1934. The 1934 statutory ceiling on autonomous operation, which caps at continuous fire until manual reload, leaves no doctrinal vocabulary for systems that source their own continuation through external compute markets, cryptocurrency wallets, and contracted services. This Article develops the statutory architecture, the post-Cargill application landscape, the application of the test to AI deployment, the doctrinal vacuum beyond the 1934 ceiling, the cross-aisle convergence, the standing implications, and the regulatory architecture that the federal regulatory state has used to address autonomous operation in adjacent contexts and that supplies a structural model for addressing it here. The honest political history of conditional anti-regulation, beginning with the Mulford Act of 1967 and the National Rifle Association’s support for prohibiting open carry in response to the Black Panther Party, is addressed because it establishes the ### 92. The Compliance Impossibility: Why Ambient Biometric Devices Cannot Comply with Amended COPPA - Category: Privacy, Surveillance, and Consent - Type: Research paper - Preferred URL: https://ssrn.com/abstract=6246539 - Local PDF: https://realsafetyai.org/documents/Compliance_Impossibility.pdf - Abstract: The Federal Trade Commission’s 2025 amendments to the Children’s Online Privacy Protection Act (COPPA) Rule expanded the definition of personal information to include biometric identifiers such as facial templates, voiceprints, and gait patterns. Under COPPA, operators must obtain verifiable parental consent before collecting personal information from children under 13. This paper identifies a structural compliance impossibility affecting an entire class of consumer technology products: ambient biometric sensing devices. These devices, which include smart glasses with facial recognition, AI powered wearable recorders, smart home cameras with person identification, cashierless retail systems, and mixed reality headsets, must process biometric data from every person in their sensing environment to function. When children are present in that environment, the device must collect a child’s biometric data to determine whether the subject is a child, but that collection itself constitutes collection of children’s personal information requiring prior parental consent under amended COPPA. The device must break the law to determine whether it is breaking the law. This paper demonstrates that this paradox is not a product specific engineering problem but a category wide structural impossibility inherent to ambient biometric sensing. We survey ten product categories encompassing dozens of shipping and announced devices from companies including Meta, Amazon, Apple, Google, Samsung, and numerous startups, demonstrating that the impossibility applies universally. We further show that existing legal scholarship and regulatory guidance has not identified this cross product structural problem, treating bystander biometric collection as a product specific compliance challenge rather than a systemic regulatory paradox. With the COPPA compliance deadline of April 22, 2026, approaching, hundreds of billions of dollars in planned product launches face a legal obstacle that cannot be engineered around. Keywords: COPPA, biometric identifiers, ambient sensing, facial recognition, children’s privacy, regulatory compliance, smart glasses, wearable AI, structural impossibility, consent paradox ## Contact and collaboration Use the Contact page for collaboration, education partnerships, research discussion, or harm database contributions: https://realsafetyai.org/contact