After reviewing 137 (109 directly referencing AI) peer-reviewed accessibility research papers published between 2021 and 2025 — from the ACM SIGACCESS Conference on Computers and Accessibility (ASSETS) and the International Web for All Conference (W4A) — a clear picture emerges of a field at an inflection point. Artificial intelligence is simultaneously the most promising tool and the most concerning threat in digital accessibility today.
The Scale of the Problem Has Not Changed
Let’s start with a sobering baseline. The WebAIM Million project, which annually audits the top one million home pages, found in its most recent analysis that 96.3% of home pages had detectable WCAG 2 failures (Hackett et al., 2024). The most common errors — low contrast text, missing alt text, empty links, missing form labels — are the same ones that have topped this list for years.
In scholarly publishing, the picture is equally bleak. An analysis of PDF accessibility across major publishers found that only 3.2% of tested PDFs satisfied basic accessibility criteria, while 74.9% failed every criterion tested (Huh et al., 2024). And when researchers examined accessibility overlays — the widgets that promise automated fixes — only 3 of 34 blind respondents who had used overlay features rated them “very effective,” while 16 rated them “not at all effective” (Alharbi et al., 2024).
Against this backdrop of persistent failure, AI enters the conversation.
AI as Accessibility Tool: Real Progress, Real Limitations
The most exciting applications of AI in accessibility are tools that do things that were previously impossible or prohibitively expensive.
Computer vision is transforming navigation and object recognition. Teachable object recognizers, tested with 15 blind and low vision participants through CNIB, achieved an 83% success rate in helping users locate objects in real-world environments (Morrison et al., 2023). WheelPoser uses just four inertial measurement units to estimate full body poses for wheelchair users with a threefold improvement over prior methods — enabling activity recognition and health monitoring applications that previously excluded this population (Li et al., 2024).
AI-generated image descriptions are improving, but remain unreliable. A Chrome extension using GPT-4V to generate context-aware descriptions — informed by surrounding text, heading hierarchy, and the image’s role on the page — outperformed generic descriptions across all quality measures in a study with blind and low vision users (Gubbi Mohanbabu & Pavel, 2024). But hallucinations persisted: the AI fabricated visual details and sometimes parroted nearby text rather than describing what was actually in the image. Compounding this, users tend to over-trust a single confident AI description. Chen et al. (2025) found that presenting variations from multiple models increased users’ ability to spot unreliable claims by 4.9 times — showing users where descriptions disagree may matter as much as improving accuracy.
Sign language technology is advancing, but data collection is the bottleneck. SignStreamNet set a new benchmark for streaming sign language video-to-text translation (Ahmed, 2025), yet building these systems depends on video data that Deaf signers are reluctant to share — 93% reported privacy concerns (Bragg et al., 2020). Filters like cel shading and avatar replacement nearly doubled willingness to contribute publicly (from 36% to 56%), and models trained on filtered data matched unfiltered performance. Respecting privacy and building better technology need not be in tension.
Writing and communication tools are finding their niche. Fluent, an AI writing tool for people who stutter, learns which words trigger disfluency for each individual — reaching over 80% accuracy after just 20 interactions — and suggests real-time synonyms to help users write fluently (Ghai & Mueller, 2021). TwIPS uses large language models to help autistic users decode ambiguous text messages — participants described the relief of having an on-demand “translator” for social nuance they might otherwise miss (Ramesh et al., 2024).
Screen magnifier users benefit from intelligent content restructuring. TableView, a browser extension that extracts web data records into compact tables, reduced task completion times by 72.9% and cognitive workload by 65.9% for low-vision screen magnifier users trying to compare items like job listings or flight results (Lee et al., 2020).
The LLM Revolution: Promise and Peril in Equal Measure
Large language models are rapidly becoming the most discussed technology in accessibility research. Of the 137 AI-related papers reviewed, 23 specifically address large language models, almost all from 2024-2025. The findings are decidedly mixed.
LLMs can detect and fix accessibility violations — but not as well as traditional tools. No LLM tested exceeded 38% of the violations found by conventional automated scanners for mobile apps (Alotaibi et al., 2025). Where LLMs show distinct value is in semantic violations that rule-based tools cannot assess — whether alt text actually conveys an image’s meaning, or whether reading order logically follows content. AccessGuru, combining GPT-4 with traditional scanners, achieved a 96% decrease in semantic violations (Fathallah et al., 2025). For PDF evaluation, a proposed three-tiered approach — automated checkers for structural issues, LLMs for semantic assessment, human experts for conflicts — reflects the emerging consensus that LLMs complement rather than replace existing tools (Kumar et al., 2025). On the fix side, results are mixed: ChatGPT fixed 70% of its own violations but achieved 0% on certain heading and form labelling issues (Aljedaani et al., 2024), and LLMs sometimes “fixed” contrast problems by simply turning everything black and white.
But LLMs generate inaccessible code by default. A systematic evaluation found that 84% of ChatGPT-generated websites contained accessibility violations, with issues overwhelmingly concentrated under WCAG’s Perceivable principle (Aljedaani et al., 2024). When AI coding assistants like Copilot are used in development, they create a “double-edged sword” — consistently generating placeholder alt attributes (good) while simultaneously introducing keyboard traps and failing to provide accessible error handling (bad) (Pandey et al., 2024).
Explicit accessibility prompting helps dramatically — but is not sufficient. Accessibility-oriented prompts reduced violation rates in LLM-generated code from 58% to 19% and decreased severity scores from 1.53 to 0.30 (Gurita & Vatavu, 2025). Landmark roles jumped from 11% to 94%, heading hierarchy from 33% to 89%, and skip links from 0% to 100%. But dynamic interaction patterns, state management, and sensory feedback remained problematic even with the best prompts.
GenAI can restructure entire webpages for accessibility. Research from Yu et al. (2025) showed that using generative AI to restructure cluttered e-commerce pages for screen reader users reduced detectable WCAG violations from 16 to as few as 1 on Amazon product pages, while preserving content integrity. This goes beyond fixing individual elements to fundamentally rethinking how pages are organized for non-visual access.
AI Fairness and Disability: The Overlooked Dimension
Perhaps the most important thread in the research is the growing body of work on how AI systems can harm people with disabilities — and how the broader AI fairness community has largely overlooked disability.
The landmark 2020 paper by Guo et al., “Toward Fairness in AI for People with Disabilities: A Research Roadmap,” identified five categories of potential harm: quality of service degradation, harms of allocation, denigration, stereotyping, and over- or under-surveillance. Voice-activated devices that cannot recognize atypical speech. Hiring systems that filter out disabled applicants. Search results that reinforce disability stereotypes. Each category maps to real, documented consequences.
By 2025, Venkatasubramanian et al. built a comprehensive taxonomy of AI harms against disabled people, analyzing real-world incidents. The most common harm — denying vital resources — accounted for 37 documented incidents, spanning welfare and social services, technology access, and employment. Medicaid resource allocation models were found arbitrary and non-transparent in multiple U.S. states. Employment screening tools systematically disadvantaged applicants with disabilities.
The concept of “dignity of risk,” introduced by Tanis and Lewis (2020), reveals a troubling pattern: many people with cognitive disabilities are responding to AI risks by withdrawing entirely — refusing to use online job applications, declining to share data, avoiding AI-enabled services. While this is an understandable response, it comes at the cost of foregone benefits and further marginalization.
LLMs encode and reproduce disability bias. Research examining age bias found that LLMs portrayed older groups negatively 105 times versus positively only 70 times for technology proficiency stereotypes alone (Strom et al., 2025). Studies of ability-bias found distinct stereotypical associations — autism linked to “difficulty expressing emotions,” visual impairment to “dependence” — that were then reproduced in LLM outputs (Cheticamp et al., 2025).
Dataset bias compounds the problem. A meta-analysis of accessibility datasets found that only 38% can be directly downloaded, the median contributor count is just 20 people, and demographic representation is uneven — with some disability communities almost entirely absent from the training data used to build tools meant to serve them (Kamikubo et al., 2022). When only 42% of datasets even have clear sharing practices (Kamikubo et al., 2021), the foundation for equitable AI is shaky.
How Blind Users Actually Experience AI Errors
One of the most illuminating studies examines how blind people verify and contest AI errors in visual assistance tools (Alharbi et al., 2024). The findings challenge the narrative that AI simply “helps” disabled users:
- AI frequently failed with complex document layouts, producing scrambled or incomplete text.
- It struggled with non-English languages, particularly Arabic.
- It misidentified cultural artifacts and foods from non-Western cultures.
- And most critically, blind users have limited ability to independently verify whether AI outputs are correct.
Most participants preferred to independently verify results rather than rely on sighted assistance — reflecting a deep desire for autonomy. But this creates what the researchers call a “misfitting” dynamic: users must develop elaborate workaround strategies for technology that is supposed to serve them.
A parallel finding from the autoethnographic work of Glazko et al. (2023): generative AI tools were most useful for low-stakes, easily verifiable tasks — generating drafts, suggesting color palettes, extracting information from dense text. For high-stakes tasks where errors matter, a “verification bottleneck” emerged: the same accessibility barriers that motivated users to seek AI help also made it difficult to check AI outputs for errors.
The Neurodivergent AI Experience
An emerging research thread explores how neurodivergent people use AI tools — and the results complicate simplistic narratives about AI as assistive technology.
Neurodivergent students found generative AI genuinely helpful for managing academic workload but reported significant tensions: challenges integrating AI into established routines, limited AI literacy surfacing as a barrier, and anxiety about whether AI use constitutes “cheating” (Jamshed et al., 2025). Students with anxiety-related conditions like OCD were particularly resistant to disrupting tried-and-tested workflows.
A provocative autoethnographic study by Moore (2025) found that generative AI tools function as effective executive function support not because they were designed as assistive technology, but because their interaction patterns — conversational scaffolding, step-by-step guidance, patient repetition — coincidentally align with what people with executive dysfunction need. The irony: obtaining ADHD medication in the U.S. requires navigating monthly prescription renewals, prior authorisations, and pharmacy phone calls that demand the very executive functions the medication treats. AI chatbots accidentally became more cognitively accessible than the healthcare system designed to help.
For autistic job seekers, LLM-based support tools introduced new problems: fabricating qualifications, adding proficiencies users did not possess, and stripping authentic communication styles from resumes and cover letters (Haroon et al., 2025). The tools optimized for neurotypical communication norms rather than helping users present themselves authentically.
The Data Challenge Remains Foundational
Across every AI application in accessibility, the data problem is the limiting factor.
Sign language recognition still lacks sufficient training data, and the privacy concerns discussed earlier constrain collection (Bragg et al., 2020). Speech recognition for people with speech disorders faces a similar scarcity, but self-supervised learning — training on unlabelled disordered speech before fine-tuning — substantially improved accuracy while reducing the costly need for expert annotation (Takashima et al., 2024).
Disability-first dataset creation — where disabled people lead the collection process — produces fundamentally different and better data. When blind users collected their own training images for teachable object recognizers, switching from photos to videos proved essential, and the resulting models were more robust to real-world conditions (Theodorou et al., 2021).
The participatory design movement in accessibility research is accelerating, reflecting a field-wide recognition that technology built without meaningful involvement of disabled people fails in predictable ways. Research centering the Deaf community in sign language AI development (Kamikubo et al., 2025) and co-designing exercise games with children who have motor impairments (Duval et al., 2023) exemplify approaches that produce both better technology and more ethical research.
What This Means for Practitioners
After synthesizing this body of research, several conclusions are clear:
AI will not solve the accessibility crisis automatically. The web remains 96% non-compliant, and AI tools generate inaccessible code by default. AI is a tool, not a solution — it requires deliberate, informed application.
Explicit accessibility requirements matter enormously. Whether you are prompting an LLM to generate code, training a computer vision model, or designing a conversational interface, accessibility must be specified upfront. The difference between 58% and 19% violation rates is the difference between a prompt that mentions accessibility and one that does not.
AI fairness research must include disability. The field of algorithmic fairness has focused primarily on race and gender. The research reviewed here demonstrates that disability creates distinct and serious AI harms — from denial of vital resources to systematic exclusion from employment — that require dedicated attention.
Nothing about us without us remains the gold standard. The most successful AI accessibility tools in this literature share a common trait: meaningful involvement of disabled people in design, data collection, and evaluation. The technology that works best is the technology built with the people who use it.
Privacy and data ethics are non-negotiable. Small, identifiable disability communities bear outsized risks from data collection. Privacy-preserving approaches are not just ethically necessary — they may actually produce more data and better models by increasing willingness to participate.
These 137 AI-focused papers span five years of research, but the most consequential work is happening now. AI is reshaping what is possible in digital accessibility. The question is whether we will use it to reinforce existing barriers — or to finally dismantle them.