When AI Policy Becomes an Accessibility Barrier

General-purpose AI may function as assistive technology when a person with a disability uses it to increase, maintain, or improve a functional capability. The companion paper names that function Artificially Intelligent Disability Assistance (AIDA) and offers a three-prong functional test for when a particular use qualifies, paired with a separate governance screen. This paper asks what follows once it does. An institution's blanket restriction, mandatory disclosure, distribution penalty, discipline, or detection rule may then implicate existing disability-law obligations, even when the AI policy is facially neutral. The paper poses a question. If general-purpose AI is a disability tool, is what institutions are doing to restrict it unethical at best, and illegal at worst? The work is to find which applies where, measured against the limits the law already keeps. The problem it examines is AI policy that restricts or penalizes the people who use AI. That covers anti-AI rules that stop the use outright, and wider rules that penalize it through disclosure mandates, labels, reduced distribution, discipline, or detection. AI Governance, which is human oversight and accountability, is the remedy. A requirement to admit or disclose assistive use may itself breach ethical duties or disability protections. It sets out the reasons institutions give for restricting AI and classifies six types of policy risk. It maps which United States statutes reach which actors, works three gray-area hypotheticals, and proposes a seven-gate screen for accommodation exposure. It compares the European Union, where the Article 50 Guidelines exempt meaning-preserving assistive communication while marking the summarizing and restructuring that cognitive disabilities often rely on. It closes with a model accommodation clause that generalizes practice already in place at the Open University and at a United Kingdom government department. Existing law, applied actor by actor, decides whether any duty follows. The register throughout is potential exposure.