Real Safety AI Foundation / Research

Disability, Accessibility, and Civil Rights

The Assistive Exemption: Content Provenance Marking and the Compelled Disclosure of Disability

Publisher: Real Safety AI Foundation

Working draft. Not peer reviewed.

Written
August 2026
Version
v0.8
Pages
35

Abstract

Article 50(2) of the EU Artificial Intelligence Act requires providers of generative systems to mark outputs in a machine-readable format so that they are detectable as artificially generated. The obligation carries an express carve-out: it does not apply where the system performs an assistive function for standard editing. The Commission's July 2026 Guidelines define that exception to cover grammar correction, spellchecking, minor stylistic polishing, machine translation, and, expressly, content that only transforms authentic human input through assistive technologies allowing persons with disabilities to communicate. They also confirm that the exception applies per operation rather than per system. So the drafters did consider disabled users. They considered one kind. The carve-out reaches assistive technology that lets a person communicate without altering semantics; it does not reach assistive technology that helps a person compose or organize, and the same Guidelines place summarization and substantive rewriting back inside the marking duty. The line runs through the middle of the disabled population. The first implementations went further still. In August 2026 Anthropic announced that its models mark generated text worldwide, and its own documentation concedes that a detected mark shows only that text may have been processed by the model, naming proofreading, translation, and summarization as uses it cannot distinguish. Two of those three are inside the Commission's exception. The mark attaches anyway, applies wherever the product is offered rather than to output used in the Union, and reaches personal use the Act excludes from scope. The Code of Practice then decides who reads it. It concedes in operative text that watermark detection for free-form text has lower reliability and may produce misleading or low-confidence results, and it nonetheless guarantees access to a named class of expert users that includes educational and research organisations. It requires that a detection result be downloadable as a digitally signed artifact with a content hash and timestamp. It contains no provision entitling the person whose text was queried to notice, to the receipt, or to any route to contest the result. This Article makes three claims. First, that accessibility enters this regime in one direction only: it constrains how a notice or a detection result is delivered to a reader, never who is marked, and the Guidelines state that Article 50 imposes no distinct accessibility requirements of its own. Second, that the resulting position is a compelled disclosure. The mark says nothing about disability; it says machine involvement, and where an institution restricts that involvement the only defense available to an accommodated author is the accommodation. A facially neutral rule answerable only by disclosing protected status is a disability related inquiry in substance, and existing doctrine already says so. Third, that no doctrine reaches the injury when it occurs. The provider that creates the signal owes the marked person nothing; the institution that acts on it holds a guaranteed right of access and did not create it. Accommodation law is built for a dyad. Provenance marking makes a triad whose third element is a property of the artifact.

Keywords

  • EU AI Act
  • Article 50
  • content provenance
  • watermarking
  • assistive technology
  • disability disclosure
  • compelled disclosure
  • meaningful access
  • differential burden
  • academic integrity
  • AI detection
  • Section 504

Plain language slides

First slide of the plain language summary of The Assistive Exemption: Content Provenance Marking and the Compelled Disclosure of DisabilityOpen the 33-slide summary (PDF)

Suggested citation

Gilly, Travis. "The Assistive Exemption: Content Provenance Marking and the Compelled Disclosure of Disability." Real Safety AI Foundation Working Draft, August 2026. https://realsafetyai.org/research/k7me58/

Other versions

This paper is also posted on SSRN.

SSRN version

References (38)

  1. Alexander v. Choate, 469 U.S. 287 (1985).
  2. Americans with Disabilities Act of 1990, 42 U.S.C. §§12101–12213.
  3. Andre Grenier v. Cyanamid Plastics, Inc., 70 F.3d 667 (1st Cir. 1995).
  4. Anthropic. (2026). Consumer terms of service. https://www.anthropic.com/legal/consumer-terms.
  5. Anthropic. (2026). How Claude marks AI-generated content. https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content.
  6. Bobnar v. AstraZeneca Pharmaceuticals LP, 672 F. Supp. 3d 475 (N.D. Ohio 2023).
  7. Carparts Distribution Center, Inc. v. Automotive Wholesaler’s Association of New England, Inc., 37 F.3d 12 (1st Cir. 1994).
  8. Cech, E. A. (2023). Engineering ableism: The exclusion and devaluation of engineering students and professionals with physical disabilities and chronic and mental illness. J. Eng’g Educ., 112, 462. https://doi.org/10.1002/jee.20522.
  9. Conroy v. New York State Department of Correctional Services, 333 F.3d 88 (2d Cir. 2003).
  10. Cummings v. Premier Rehab Keller, P.L.L.C., 596 U.S. 212 (2022).
  11. Dathathri, S., et al. (2024). Scalable watermarking for identifying large language model outputs. Nature, 634, 818.
  12. Enyart v. National Conference of Bar Examiners, Inc., 630 F.3d 1153 (9th Cir. 2011).
  13. European Commission. (2026). Guidelines on transparency obligations for providers and deployers of certain AI systems. https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content.
  14. European Commission. (2026). Transparency obligations under Article 50 of the AI Act. https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act.
  15. Federal Trade Commission Act, 15 U.S.C. §45.
  16. Gaba, P., & De Cristofaro, E. (2026). Group-differentiated discourse on generative AI in high school education. arXiv:2603.24972.
  17. Gegg-Harrison, W., & Quarterman, C. (2024). AI detection’s high false positive rates and the psychological and material impacts on students. In Academic integrity in the age of artificial intelligence (p. 199). IGI Global. https://doi.org/10.4018/979-8-3693-0240-8.ch011.
  18. Ibrahim, K. H. S., AlOtaibi, D., & Sibai, F. N. (2025). The robustness of AI-classifiers in the face of AI-assisted plagiarism. Int’l J. Computer-Assisted Language Learning & Teaching, 15, 1. https://doi.org/10.4018/ijcallt.375428.
  19. Illinois Consumer Fraud and Deceptive Business Practices Act, 815 ILCS 505.
  20. Kirchenbauer, J., et al. (2023). A watermark for large language models. In Proceedings of the 40th International Conference on Machine Learning.
  21. Kroll v. White Lake Ambulance Authority, 691 F.3d 809 (6th Cir. 2012).
  22. Liang, W., et al. (2023). GPT detectors are biased against non-native English writers. Patterns, 4, 100779. https://doi.org/10.1016/j.patter.2023.100779.
  23. Nondiscrimination on the Basis of Disability by Public Accommodations, 28 C.F.R. §36.309.
  24. Payan v. Los Angeles Community College District, 11 F.4th 729 (9th Cir. 2021).
  25. Payan v. Los Angeles Community College District, 169 F.4th 971 (9th Cir. 2026).
  26. Regulation (EU) 2024/1689 (Artificial Intelligence Act), 2024 O.J. (L 1689).
  27. Rehabilitation Act of 1973 §504, 29 U.S.C. §794.
  28. Robinson v. Toyota Motor Credit Corp., 201 Ill. 2d 403 (Ill. 2002).
  29. Robles v. Domino’s Pizza, LLC, 913 F.3d 898 (9th Cir. 2019).
  30. Sadasivan, V. S., et al. (2023). Can AI-generated text be reliably detected? arXiv:2303.11156.
  31. Santuzzi, A. M., Waltz, P. R., Finkelstein, L. M., & Rupp, D. E. (2014). Invisible disabilities: Unique challenges for employees and organizations. Indus. & Org. Psychol., 7, 204. https://doi.org/10.1111/iops.12134.
  32. Shekhar, A. R., et al. (2026). DoPE: Decoy oriented perturbation encapsulation. arXiv:2601.12505.
  33. Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., et al. (2023). Testing of detection tools for AI-generated text. Int’l J. for Educ. Integrity, 19, art. 26. https://doi.org/10.1007/s40979-023-00146-z.
  34. Coley, M. (2023). Guidance on AI detection and why we’re disabling Turnitin’s AI detector. Vanderbilt University Brightspace. https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/.
  35. Elkhatat, A. M., Elsaid, K., & Almeer, S. (2023). Evaluating the efficacy of AI content detection tools in differentiating between human and AI-generated text. Int’l J. for Educ. Integrity, 19, art. 17. https://doi.org/10.1007/s40979-023-00140-5.
  36. European Commission. (2026). Code of practice on transparency of AI-generated content. https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content.
  37. European Commission. (2026). Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act. https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems.
  38. Wechuli, Y. (2024). Passing as able-bodyminded, disabled, or supercrip: Rethinking impression management strategies through a disability lens. Symbolic Interaction, 48, 293. https://doi.org/10.1002/symb.711.

All research