Digital accessibility compliance in U.S. higher education has operated as a persistent and systemically unresolved institutional challenge since the passage of the Americans with Disabilities Act in 1990 (ADA, 1990). Despite decades of federal guidance, litigation, and voluntary resolution agreements, the sector has demonstrated a durable pattern of prioritizing policy adoption over the behavioral and infrastructural change that compliance requires (Epshteyn, 2019; Seale et al., 2015). The April 2024 Department of Justice final rule under Title II of the ADA, establishing WCAG 2.1 Level AA as the enforceable standard for web and digital content produced by public entities, including public institutions of higher education, introduced the most significant regulatory development in the sector’s accessibility compliance history, with phased compliance deadlines beginning in 2026 (U.S. Department of Justice, 2024). Simultaneously, the rapid and largely ungoverned adoption of generative AI tools across institutional functions has introduced a compounding variable that existing compliance frameworks are not designed to address.
AI-driven tools – from automated captioning to document remediation, have generated legitimate discussion about whether the technology might finally help close a compliance gap the sector has never resolved (Burgstahler, 2020). But, can Ai help us with genuine compliance progress or is it an emerging liability hiding behind promising technology?
Where AI Shows Real Promise and Where That Promise Ends
AI-driven tools have produced measurable improvements in specific accessibility functions. Automatic speech recognition has meaningfully raised real-time captioning accuracy, with error rates declining significantly as machine learning models have matured (Microsoft, 2023). AI-powered document remediation tools have reduced the time required to tag PDFs and correct reading order in high-volume content environments. For institutions managing large digital content libraries with shrinking budgets and limited accessibility staffing, these are not trivial gains.
But these gains are narrow, function-specific, and conditional on human oversight to meet WCAG 2.1 standards. On captioning, the 2023 State of Automatic Speech Recognition — a widely cited industry benchmark — found that even the most accurate ASR solutions on the market achieved no more than 93% accuracy under ideal audio conditions, with accuracy dropping as low as 57.5% when background noise, specialized terminology, or non-native speakers were present (3Play Media, 2023). Those conditions are routine, not exceptional, in higher education classrooms. The W3C Web Accessibility Initiative is unambiguous on what that means for compliance: ASR-generated captions do not meet accessibility requirements unless they can be confirmed to be fully accurate (Lawton Henry et al., 2022). On alt text, a 2024 study examining screen reader user perspectives on AI-generated image descriptions found that automated outputs frequently failed to capture the provenance, context, or intended meaning of images in ways that mattered to blind and low-vision users (Das et al., 2024). Reducing friction at a specific point in a compliance workflow is not the same as achieving compliance. And in the context of ADA Title II, the difference between the two is legally consequential.
Accessibility Checkers
Built-in LMS accessibility checkers represent another area where automated tools are doing genuinely useful institutional work. Tools such as Instructure’s Ally, integrated into Canvas and Blackboard, use rule-based automated scanning to flag accessibility issues in uploaded course content – missing alt text, untagged PDFs, improper heading structure – and provide faculty with guided remediation steps and accessibility scores at the course level (Anthology Inc., 2024). For institutions managing thousands of courses across dozens of departments, that kind of automated visibility into content-level compliance gaps is meaningful.
AI-driven tools can improve on rule-based checkers like Ally in one meaningful way – context. Rule-based checkers apply fixed criteria uniformly, so they cannot distinguish between a decorative image that WCAG says needs no alt text and an informative graphic that does. Modern AI-driven tools are beginning to make that contextual distinction, which reduces false positives and catches issues rule-based scanning misses (TestParty, 2025). But the ceiling is the important part. The research is consistent: automated tools – whether rule-based or AI-driven – detect somewhere between 25% and 40% of WCAG issues. The remaining 60-75% require human judgment. Focus order logic, whether alt text actually conveys meaningful information in context, content readability under assistive technology, and complex interactive elements all fall outside what any automated system can reliably assess (TestParty, 2025; ADACP, 2025).
That ceiling matters legally. Flagging a problem is not fixing it, and detecting 40% of violations is not conformance. Whether faculty act on accessibility flags (and how consistently) depends on the same institutional conditions that determine compliance outcomes across every other domain: clear behavioral expectations, active sponsorship, and accountability mechanisms (Harrison & Stokes, 1992; Epshteyn, 2019). The checker, whether rule-based or AI-driven, is a useful instrument but not a substitute for the implementation infrastructure required to produce consistent behavioral change.
User Companions (3d Party Overlays)
Nowhere is the gap between AI’s compliance promise and compliance reality more stark than in the proliferation of AI-powered accessibility overlays — third-party JavaScript tools marketed to institutions as automated, one-line-of-code solutions to WCAG conformance. Products such as accessiBe, UserWay, and AudioEye are all marketed as AI-driven, though the nature of that AI varies considerably across them. accessiBe claims to use AI to automatically scan and remediate accessibility violations in real time. UserWay markets itself as AI-powered but in practice uses AI primarily to detect and flag issues rather than fix them. AudioEye deploys the most substantive AI of the three — running WCAG scans, assigning compliance scores, and deploying automated fixes — supplemented by human expert review at premium tiers (TestParty, 2025; Hounder, 2026). Despite these differences, all three operate by injecting script into a website’s rendered page after it loads, and all three share the same fundamental limitation: none can remediate the underlying structural and semantic code failures that WCAG 2.1 conformance requires. The appeal to institutional procurement decision-makers is obvious. The evidence against them is unambiguous and growing.
Do They Help with Compliance?
In 2021, hundreds of accessibility professionals – including WCAG specification contributors, internal accessibility experts at companies such as Google, Microsoft, and Apple, and accessibility staff from higher education institutions including MIT, Carnegie Mellon, Syracuse, and the University of Massachusetts, signed the Overlay Fact Sheet, an open statement documenting that overlays cannot achieve full WCAG conformance and in many cases actively interfere with the assistive technologies users already rely on (Groves et al., 2021). A WebAIM survey of disabled users found that 72% rated accessibility overlays as not at all or not very effective (shouldiuseanaccessibilityoverlay.com, 2021). The National Federation of the Blind issued formal resolutions against overlay providers, returning sponsorship funds and publicly rejecting the premise that automated overlays constitute genuine accessibility (National Federation of the Blind, 2021). The most definitive regulatory statement came in April 2025, when the Federal Trade Commission finalized a $1 million order against accessiBe specifically for its AI-powered compliance claims — finding that despite marketing its tool as capable of making any website WCAG compliant, those claims were false, misleading, or unsubstantiated, and barring the company from making such representations going forward (Federal Trade Commission, 2025). The FTC action did not apply only to accessiBe. It established a legal standard against which identical claims made by UserWay, AudioEye, and similar AI-powered overlay providers are now directly measured (TestParty, 2025). The European Commission has taken an equally unambiguous position, publicly stating that accessibility overlays — whether AI-powered or not — do not constitute a valid path to WCAG conformance under the European Accessibility Act, which took full effect in June 2025 (Northwoods Digital, 2025).
The litigation data reinforces that position. According to UsableNet’s 2024 digital accessibility lawsuit report, 25% of all accessibility lawsuits filed that year — 1,023 cases — explicitly cited accessibility overlay widgets not as solutions but as barriers (UsableNet, 2024). That trend has not slowed. Over 300 additional overlay-related suits were filed in the first half of 2025 alone, with some resulting in court-mandated removal of the overlays and settlement agreements requiring full structural remediation (Northwoods Digital, 2025). Courts have consistently ruled that simulating accessibility is not the same as meeting accessibility standards. For higher education institutions operating under the DOJ’s April 2026 compliance deadline, procuring an AI-powered overlay is not a compliance strategy. The data and the regulatory record indicate it is a liability.
The Content Generation Problem
The more pressing issue is not what AI does imperfectly in accessibility-specific workflows but what generative AI is doing at scale across every other institutional workflow simultaneously.
Faculty adoption of generative AI for course content creation has grown substantially since the public availability of large language model platforms beginning in late 2022, with a significant and growing proportion of higher education faculty reporting use of these tools to produce instructional materials (Trust et al., 2023). The content those tools produce, such as PDFs, slide decks, and web-based documents, routinely lacks the structural metadata WCAG 2.1 requires: semantic heading hierarchies, tagged reading order, text alternatives for non-text content, and synchronized captions for time-based media (World Wide Web Consortium, 2018). AI-generated images arrive without meaningful alternative text. AI-generated video content enters institutional systems without captions meeting WCAG 1.2 timing and accuracy standards.
In other words, the rate at which generative AI is enabling the creation and distribution of inaccessible content inside institutional learning management systems now exceeds what any existing accessibility review or remediation process was designed to handle (Burgstahler, 2020). Institutions that spent years building remediation workflows for legacy content are now facing a real-time content generation pipeline that those workflows cannot absorb.
Institutional Procurement
The 2025 EDUCAUSE AI Landscape Study and WCET’s 2025 institutional survey document AI tool adoption across six broad functional categories in U.S. higher education (Robert & McCormack, 2025; WCET, 2025). Each carries its own accessibility and equity risk profile, and none has accessibility as a primary design criterion.
- Generative AI integrated into productivity suites institutions already own. Microsoft made Copilot available to all higher education students and faculty through Microsoft 365 in December 2023. Google added Gemini to Google Workspace for Education in May 2024 (EdTech Magazine, 2025). These tools are not being procured through a deliberate institutional decision-making process — they are arriving as features inside platforms institutions have already paid for, with accessibility implications that procurement offices did not evaluate because there was no discrete procurement decision to trigger a review.
- Intelligent tutoring systems and adaptive learning platforms – tools that adjust content, pacing, and feedback based on individual student performance in real time. Adoption is growing, driven largely by retention data that institutions find compelling (QuadC, 2025).
- AI-powered student advising and chatbot platforms, deployed for enrollment support, academic advising, and student services functions (operating largely without human oversight at the point of student contact).
- AI proctoring platforms – using facial recognition, eye tracking, and behavioral monitoring to enforce academic integrity during remote assessments. These have expanded rapidly. A 2025 Higher Education Policy Institute report found that 88% of university students are now using AI for assessments, which is accelerating institutional adoption of AI-based monitoring tools in direct response (HEPI, 2025).
- AI writing detection tools, adopted widely and quickly despite a documented and growing body of evidence questioning their reliability for diverse student populations.
- Predictive analytics platforms – early alert and student success systems that analyze behavioral, academic, and engagement data to identify and intervene with at-risk students.
Taken together, these six categories represent a substantial and accelerating institutional investment in AI tools that touch nearly every dimension of the student experience – learning, advising, assessment, and support. The AAUP’s 2025 national survey found that while 90% of institutions have introduced AI initiatives of some kind, only 20% have published governing policies for AI use, including in teaching and research (AAUP, 2025). Institutions are not making careful, bounded decisions about AI adoption. They are managing a sprawl of tools, many of which arrived without formal procurement review, in an environment where the legal floor for digital accessibility compliance has just been significantly raised.
AI in Learning and Assessment
The promise in this space is genuine. An integrative review of peer-reviewed research published between 2020 and 2024 found that AI-driven assistive technologies – including speech-to-text tools, adaptive interfaces, and context-based communication prediction, have demonstrated measurable improvements in academic engagement and performance for students with a range of disabilities in higher education (Dumitru et al., 2026). A 2024-2025 report from the Center for Democracy and Technology found that nearly six in ten teachers of students with disabilities used AI to help develop individualized accessibility and education plans (Patrick, 2025). This seems to indicate that AI, designed and deployed with intention, can reduce barriers for students who have historically faced the greatest access challenges in higher education.
The problem is that intentional design and deployment are still not the norm.
A 2025 systematic review of AI-based learning tools in higher education found that none of the 63 peer-reviewed studies reviewed addressed the needs of students with disabilities, and that AI learning tool adoption has the potential to exacerbate educational inequities by widening existing access gaps (Springer Nature, 2025). That finding reflects a structural reality: people with disabilities are underrepresented on the design and development teams building these tools, which means the datasets, interaction models, and design assumptions embedded in AI learning platforms do not reflect their experiences or needs (Every Learner Everywhere, 2025). As Dumitru et al. (2026) concluded, biases in AI algorithms can disproportionately affect students with disabilities, limiting accessibility or reinforcing inequalities rather than reducing them.
Nowhere is that dynamic more acute than in AI-powered assessment tools.
AI proctoring platforms have been found to disproportionately flag students with disabilities for suspicious behavior – particularly students with involuntary movements, tic disorders, or who rely on assistive technologies during testing (Dumitru et al., 2026). Research on online proctored systems documents that disabled students experience heightened surveillance-related anxiety, face asymmetric information-sharing practices that undermine their ability to self-advocate, and encounter systems designed without universal design principles embedded from the outset (Surveillance and Disability in Online Proctored Exams, 2024). The behavioral models these systems use to define normal test-taking behavior encode normative assumptions about how students move, look, and interact, and these assumptions are inherently discriminatory against students whose disabilities produce non-normative presentation.
AI writing detection tools introduce a parallel and equally serious equity problem. A 2026 mathematical analysis of AI text detectors found that current tools are structurally unreliable for diverse student populations, with a growing body of empirical evidence documenting that detectors disproportionately flag non-native English speakers and students whose writing patterns diverge from the normative linguistic profiles on which these tools were trained (AI Detectors Fail Diverse Student Populations, 2026). Students with learning disabilities (whose written expression may differ in structure, syntax, or fluency from neurotypical norms) are directly exposed to this risk. The institutional consequences of a false positive are significant: academic integrity proceedings, grade penalties, and reputational harm. That risk is not distributed equally across the student population.
The U.S. Department of Education’s Office for Civil Rights recognized this in November 2024, releasing guidance specifically on avoiding the discriminatory use of AI in educational settings, explicitly citing student activity monitoring software, facial and movement recognition tools, and AI-driven assessment systems as use cases carrying documented civil rights risk under the ADA, Section 504 of the Rehabilitation Act, and Title IV of the Civil Rights Act (U.S. Department of Education, Office for Civil Rights, 2024).
What's the Verdict?
So, would AI move the needle on digital accessibility compliance in higher education, or would it move the goalpost? Based on the available evidence, the more accurate answer is that AI is doing both, and institutions are not yet equipped to distinguish between the two. The discrete accessibility gains AI delivers in specific tool-based applications are real. But they are being outpaced by the volume of inaccessible content generative AI is enabling at the institutional level, compounded by the accessibility gaps in the AI tools themselves, and undermined by governance structures that are not positioned to manage either.
References
- 3Play Media. (2023). State of automatic speech recognition. 3Play Media. https://www.3playmedia.com
- ADACP. (2025). AI and accessibility: How machine learning is changing WCAG compliance. ADA Compliance Pros. https://www.adacompliancepros.com
- AI Detectors Fail Diverse Student Populations. (2026). A mathematical framing of structural detection limits. arXiv. https://arxiv.org/pdf/2603.20254
- Americans with Disabilities Act of 1990, Pub. L. No. 101-336, 104 Stat. 328 (1990).
- American Association of University Professors. (2025). Artificial intelligence and academic professions. AAUP. https://www.aaup.org
- Anthology Inc. (2024). Ally for LMS. Anthology. https://www.anthology.com
- Burgstahler, S. (2020). Creating inclusive learning opportunities in higher education: A universal design toolkit. Harvard Education Press.
- Das, M., Tariq, A., & Kane, S. K. (2024). “It’s not just about the image”: Understanding screen reader users’ perspectives on AI-generated image descriptions. ACM CHI Conference on Human Factors in Computing Systems.
- Dumitru, C., Abdulsahib, G. M., Khalaf, O. I., & Bennour, A. (2026). Integrating artificial intelligence in supporting students with disabilities in higher education: An integrative review. SAGE Journals. https://journals.sagepub.com/doi/10.1177/10554181251355428
- EdTech Magazine. (2025). How higher ed institutions are using built-in generative AI tools. https://edtechmagazine.com
- Epshteyn, E. (2019). From policy to compliance: US higher education faculty concerns over institutional digital content accessibility policies [Doctoral dissertation, Northeastern University]. ProQuest Dissertations & Theses.
- Every Learner Everywhere. (2025). Accessible AI requires involving and collaborating with people with disabilities. https://www.everylearnereverywhere.org
- Federal Trade Commission. (2025, January). FTC takes action against accessiBe for deceptive claims about web accessibility product. https://www.ftc.gov
- Gleason, C., Patkin, D., Kuber, R., & Wu, S. (2020). Disability and the AI accessibility gap. ACM Conference on Human Factors in Computing Systems.
- Groves, K., et al. (2021). Overlay fact sheet. https://overlayfactsheet.com
- Harrison, D., & Stokes, H. (1992). Diagnosing organizational culture. Pfeiffer.
- Higher Education Policy Institute. (2025). HEPI AI survey 2025. HEPI. https://www.hepi.ac.uk
- Hounder. (2026, March). accessiBe vs. AudioEye vs. UserWay vs. EqualWeb: An honest comparison for public agencies. https://hounder.co
- Lawton Henry, S., Abou-Zahra, S., & Brewer, J. (2022). Captions/subtitles. W3C Web Accessibility Initiative. https://www.w3.org/WAI
- Lazar, J., Goldstein, D., & Taylor, A. (2015). Ensuring digital accessibility through process and policy. Morgan Kaufmann.
- Microsoft. (2023). Azure cognitive services speech-to-text documentation. Microsoft. https://azure.microsoft.com
- National Federation of the Blind. (2021). Resolution on web accessibility overlays. https://www.nfb.org
- Northwoods Digital. (2025). Accessibility overlays in 2025: A shortcut companies should continue to avoid. https://www.nwsdigital.com
- Patrick, B. J. (2025). Keeping humanity at the center: Accessibility and artificial intelligence in education. Education International. https://www.ei-ie.org
- QuadC. (2025). Top trends in AI for higher education in 2025. https://www.quadc.io
- Rehabilitation Act of 1973, Pub. L. No. 93-112, 87 Stat. 355 (1973).
- Robert, J., & McCormack, M. (2025). 2025 EDUCAUSE AI landscape study: Into the digital divide. EDUCAUSE. https://www.educause.edu
- Seale, J., Draffan, E. A., & Wald, M. (2015). Exploring disabled learners’ experiences of e-learning. British Journal of Educational Technology, 41(6), 986–1001. https://doi.org/10.1111/j.1467-8535.2010.01078.x
- shouldiuseanaccessibilityoverlay.com. (2021). Survey results: Disabled users and accessibility overlays. https://shouldiuseanaccessibilityoverlay.com
- Springer Nature. (2025). Design and assessment of AI-based learning tools in higher education: A systematic review. International Journal of Educational Technology in Higher Education. https://link.springer.com
- Surveillance and Disability in Online Proctored Exams. (2024). Student perspectives and design implications. arXiv. https://arxiv.org/pdf/2511.10826
- TestParty. (2025). UserWay vs. accessiBe vs. AudioEye: Which overlay should you choose? https://testparty.ai
- Trust, T., Whalen, J., & Mouza, C. (2023). Editorial: ChatGPT: Challenges, opportunities, and implications for teacher education. Contemporary Issues in Technology and Teacher Education, 23(1).
- UsableNet. (2024). 2024 digital accessibility lawsuit report. https://usablenet.com
- U.S. Department of Education, Office for Civil Rights. (2024). Avoiding the discriminatory use of artificial intelligence. https://www.ed.gov
- U.S. Department of Justice. (2024). Nondiscrimination on the basis of disability; accessibility of web information and services of state and local government entities. Federal Register, 89(79), 31320–31571.
- WCET. (2025). WCET annual report: AI in higher education. WCET. https://wcet.wiche.edu
- World Wide Web Consortium. (2018). Web content accessibility guidelines (WCAG) 2.1. https://www.w3.org/TR/WCAG21/
