- From: Janina Sajka <janina@a11y.nyc>
- Date: Wed, 8 Jul 2026 05:31:27 -0400
- To: public-rqtf@w3.org
Dear All: With assistance of my LLM du jour we have the following list of issues and concerns regarding AI as logged in our minutes this year: # AI-Related Concerns Identified by the RQTF in 2026 A consolidated, itemized list of the accessibility concerns raised in the Research Questions Task Force meetings during 2026. Each item is a distinct issue. 1. AI output is plausible but unverifiable, with no reliable way for a user to confirm that it is correct. 2. The same image submitted to a generative model produces different alternative-text descriptions each time, with no way to know which, if any, is accurate. 3. It is unclear whether ordinary reasoning is enough to validate AI output, or whether genuine subject-matter expertise is required. 4. People who lack the relevant expertise cannot judge whether AI output is correct, creating an equity gap. 5. Using AI tools well depends on strong reasoning and analytical skill, so people with weaker skills in these areas are more likely to accept false information. 6. AI mistakes are frequently not obvious, so users must adopt an attitude of critical acceptance and continuous vigilance. 7. LLM-based speech recognition can confabulate plausible text that was never actually spoken, seriously disadvantaging people who depend on captions or transcripts. 8. Automatically generated alt text is hard to verify, and there is no easy human-in-the-loop backup when it is wrong. 9. Defining the scope of the work is difficult: whether to address neural-network machine learning narrowly or "AI" broadly, and how to signal which claims apply to which kind of system. 10. Rule-based, deterministic AI and neural-network machine learning behave very differently, and the variable, hard-to-predict output of the latter is what matters most for disability. 11. Discussion of AI risks getting lost in background and definitions rather than telling users what the technology does for them and how it affects them. 12. When an agent acts on an inaccessible website on a user's behalf, the same barriers that prompted the delegation also prevent the user from verifying or correcting what the agent did. 13. Compartmentalizing agents for security, by restricting their access to file systems, API keys, and root privileges, has direct accessibility consequences. 14. Security restrictions and accessibility needs can pull in opposite directions. 15. Autonomous agents can flood developers with low-quality, AI-generated bug reports and pull requests, creating a denial-of-service-like burden. 16. If users interact with the web only through agents, source websites may shrink to API-only endpoints, eroding the large accessible surface of today's web. 17. Control over branding and presentation shifts into a negotiation between content providers and agent providers rather than remaining with the user. 18. It is unclear what is actually being licensed, in copyright and intellectual-property terms, when agents mediate and repackage content. 19. In an agent-mediated web, remediation becomes the preferred accessibility strategy, with human intervention as a costly and limited fallback. 20. There is no clear way to establish trust that synthesized or agent-delivered content is correct. 21. The large data centers behind AI carry significant environmental and economic costs. 22. As ML-optimized hardware becomes standard and, increasingly, necessary, users face new and uneven hardware-purchasing burdens. 23. Beneficial "non-interactive" AI features operate without the user noticing, making them hard to evaluate or trust. 24. Privacy considerations complicate the trade-off between local, on-device processing and cloud processing. 25. Machine-learning text-to-speech remains too slow and high-latency for interactive screen-reader use. 26. Synthetic-speech audio description is rated poorly on emotional relevance even when listeners do not identify it as synthetic. 27. Guidance calling for neutral audio-description delivery conflicts with what listeners actually find satisfying. 28. Emerging external AI standards must be tracked, cited, and reconciled with existing accessibility standards. 29. Automated decision-making can affect disability rights and produce representational harms. 30. The governance of AI systems has accessibility implications that are not yet well addressed. 31. There is no reliable mechanism to indicate whether content on a page was generated by AI. 32. AI models appear biased against producing accessible code, because a large share of their training data is itself inaccessible. 33. It is uncertain whether models can be tuned, or their training data evaluated against WCAG, to overcome this bias, and doing so is genuinely difficult given how neural networks work. 34. AI-generated code enters the web at scale without accessibility built in. 35. It is unresolved whether AI methods proven for finding security defects can be applied to detect and remediate accessibility defects. 36. As agents defeat Turing tests, CAPTCHA is likely to return as a barrier, and the human-verification challenge itself raises accessibility problems. 37. Audio CAPTCHAs can be effectively unsolvable, for example whispered synthetic digits placed over a difficult noise floor. 38. Machine-learning-based identity and age verification remains an ongoing accessibility challenge. 39. Technology carries the biases of the society that produces it, so AI can inherit and amplify existing inequities. 40. AI may displace high-skilled labor with low-skilled, low-agency work. 41. "AI slop" and model collapse, where AI is trained on AI-generated content, degrade quality over time. 42. Access to the benefits of AI is not equitably distributed. 43. AI represents a change to human life and consciousness large enough that the accessibility community needs to name clearly what is coming rather than treat the technology as simply good or bad. -- Janina Sajka (she/her) https://linkedin.com/in/jsajka Digital Accessibility Standards Lead, Level Access https://www.levelaccess.com/ W3C Accessible Platform Architectures (APA) CoChair, http://www.w3.org/wai/apa Linux Foundation Fellow https://www.linuxfoundation.org/about/leadership#team-bio-janina-sajka
Received on Wednesday, 8 July 2026 09:31:32 UTC