Research

This page points to the primary sources behind the rest of the site. It is a starting set rather than a complete bibliography, and each item is something you can cite or learn from directly. Where a source shaped a specific page, that page links to it as well.

Peer-reviewed research

• Glazko and colleagues, an autoethnographic case study of generative AI for accessibility. Disabled researchers document the real benefits and harms of generative AI in their own accessibility work.
• Mason Marks, Algorithmic Disability Discrimination. A legal analysis of how automated systems infer and act on disability, often invisibly.
• Hicks, Humphries, and Slater, “ChatGPT is bullshit”. Argues that model falsehoods are indifference to truth rather than occasional error, which makes confident wrongness structural.

Cross-cultural AI design

• Luan Haoyue and Cho, Factors influencing intention to engage in human-chatbot interaction. A study of the Gov.sg chatbot (N = 304) finding that high-context users respond primarily to social presence while low-context users respond primarily to performance, showing that a single interaction model does different work for different cultural orientations.
• Naidoo and Chadha, Culturally responsive AI chatbots: From framework to field evidence. The CRAIF-C framework, tested across four linked studies over twelve months, showing that cultural communication style, narrative structure, and tonal pattern consistently produce higher trust and satisfaction when matched to the user’s cultural orientation.
• Arambewela, The Impact of Cultural Context on User Trust in AI-Powered Chatbots for Everyday Mental Health Support. A Sweden-versus-Sri-Lanka comparison showing how collectivist norms and face culture shape what users need from a mental health chatbot, with implications for when cultural misalignment becomes a harm rather than a friction problem.
• Müller and colleagues, Does Culture Matter for the Design of Chatbots Promoting Blood Donation Behaviour? A study across four countries finding that collectivist framing outperforms individualist framing even for users who self-describe as individualist, when the product is asking them to do something for others.
• More, Consumer Trust in AI-Powered Customer Service: A Cross-Cultural Study. Research across the USA, Germany, and India finding that efficiency and availability are universal trust predictors, while warmth matters more in collectivist cultures and technical accuracy matters more in individualistic ones.
• Alsswey and Al-Samarraie, The role of Hofstede’s cultural dimensions in the design of user interface: The case of Arabic. Demonstrates how cultural dimensions predict concrete interface preferences, including that icon-based communication outperforms text-heavy design for Arab users.

Synthetic users and AI usability testing

Synthetic users are AI personas meant to stand in for real testers. They matter for accessibility because disabled people are the users a model trained on web data represents least well, so a simulated panel tends to miss them and the real testers it would replace are often the ones a product most needs. These sources cover both the pitch and the evidence against it.

• Lu and colleagues at Amazon and Northeastern, an LLM agent-based usability testing framework for web design. Builds a system that generates thousands of simulated users to walk through a website, with one module driving the agents and another connecting them to a real browser. The authors frame it as a way to pilot and refine a study design before running it with real people rather than as a replacement, and their evaluation with five UX researchers welcomed the idea while raising concerns about leaning on simulated users.
• Researchers at UXtweak and the Slovak University of Technology, a systematic review of synthetic participants generated by large language models. A review of 182 studies that sorts the failures of LLM-generated participants into four kinds, cognitive misalignments, distortions, misleading believability, and overfitting or contamination. It concludes that synthetic participants cannot replace human ones, and that prompt techniques like few-shot examples or chain-of-thought reasoning give only marginal improvement because the underlying problems stay the same.
• Microsoft Research, OmniParser v2, turning a screenshot into structured elements an agent can act on. This is the perception layer an agent needs before it can use a real interface, and it is near the current state of the art for finding clickable regions on screen. On a benchmark of high-resolution screens with small icons it still reaches only 39.6 percent average accuracy paired with GPT-4o, which caps how reliable any AI that drives a real interface can be today.

Standards and policy

• The W3C AI Accessibility draft maps AI across authoring tools, interfaces, evaluation, and assistive technology.
• The W3C report on AI and the web proposes standardization steps for AI’s impact, including the problem of under-represented data.
• Accessibility Standards Canada’s CAN-ASC-6.2:2025 is an outcomes-based standard for accessible and equitable AI systems.
• The European Disability Forum’s guide to monitoring the EU AI Act tracks implementation from a disability-rights view. See also the Guidelines section.
• The US Access Board’s hearings on AI and disability signal where US policy is heading.

Industry and tooling evaluations

• The Microsoft accessibility LLM evaluation report measures how often generated markup is accessible, from 12 percent unguided to 86 percent with a dedicated skill.
• Karl Groves on remediation and AI audits quantifies where AI assistance breaks down.
• The Deque accessibility coverage report is the usual source for the automation ceiling, the share of issues automated tools can find.
• The WebAIM Million annual analysis of the top one million home pages, whose 2026 edition found accessibility errors rising for the first time in six years, attributed in part to automated and AI-assisted coding.
• David Mello on what axe and Lighthouse miss is a practitioner account of the gap between a clean automated run and real conformance, and which missed checks a behavioral test can still cover. His walkthrough of testing AI chatbots and agents covers non-deterministic output, guardrails, bias, and the accessibility checks a chat interface needs in both its open and closed states.
• Adrian Roselli on Google’s Modern Web Guidance tested Google’s AI coding guidance against the accordion prompt from its own homepage and found the result was not WCAG conformant, ignored the ARIA Authoring Practices Guide accordion pattern, and broke in Firefox. The takeaway is that a deterministic, vetted pattern library beats trusting generated code.

Community writing

• Scott O’Hara on live regions and the output element, the technical basis for announcing streamed output.
• James Scholes on ARIA and experiential design, on designing the screen reader experience rather than delegating it to ARIA.
• Adrian Roselli on ARIA, AI, and SEO, on why pushing ARIA for machines harms accessibility.
• Eric Bailey on why alt text is context dependent and on instructing an LLM with the ARIA Authoring Practices Guide.
• Hidde de Vries on teaching LLMs or organizations, on where the real bottleneck sits.
• Jakob Nielsen on accessibility and generative UI, included as the contrarian view this site argues against.
• Anna E. Cook on AI doesn’t fix accessible systems, it depends on them, arguing that AI inherits and reproduces broken structure rather than repairing it, so accessible and parseable systems are the precondition for AI to work.
• Anna E. Cook on governing accessible design systems in the AI era, a six-question framework that treats AI accessibility as a governance problem, including who owns AI output, which properties are protected, and why a model should never override an explicit user choice.
• Anna E. Cook on accessible design as digital infrastructure, on why accessibility is foundational rather than a finishing touch and how it erodes quietly when maintenance stops.
• Anna E. Cook on designing accessibility for real use, not dashboards, on why a high accessibility score is a floor rather than proof, and how scoreboards create accessibility theater.

Ongoing work and community

• The W3C AI Accessibility issues tracker is where the draft is being worked out in the open.
• The European Disability Forum’s Empowered by AI project builds AI literacy among disabled people so they can shape these systems.
• Penn State on learned disability bias in AI models is concrete evidence for the bias discussion.

Audience: Accessibility Specialist

These are useful when WCAG runs out. For the parts of AI behavior with no success criterion, such as communicating uncertainty, you can anchor findings in the peer-reviewed work and the standards drafts here rather than in WCAG alone. The industry evaluations are also handy for setting expectations with a team, since a figure like 12 percent of unguided output passing checks lands harder than a general warning.

Audience: Self-advocate

The first-person and community sources here are worth reading directly, because they describe what these tools are actually like to use rather than how they are meant to work. The autoethnographic study and the community blogs in particular center disabled people’s own accounts, which is the kind of evidence that tends to be missing from vendor material.

Further reading

• The Foundations section for how these sources fit together.
• The Case studies page for the same evidence as concrete situations.