AI & accessibility
Who is the AI talking to?
Voice assistants and chatbots promise access for everyone. The standards that exist were written before large language models, and almost nobody is designing AI conversation around memory, attention and processing. This is what we found, and what teams can do now.
What we found
Each finding carries a label: Standard (written in a standard or law), Research (peer-reviewed), Preprint (not yet peer-reviewed).
- Standard
No standard covers LLM conversation yet.
The main W3C document on natural language interfaces (NAUR) dates from September 2022 and never mentions large language models. W3C's 2026 draft on machine learning and generative AI has no section on conversation or on cognitive accessibility.
- Research
Atypical speech gets the most investment.
The Speech Accessibility Project (University of Illinois with Amazon, Apple, Google, Meta and Microsoft) had about 2,000 participants by September 2025. It is English-only so far.
- Research
Cognitive accessibility gets far less.
Studies with people with intellectual disabilities and dementia find assistants that time out on pauses, misrecognise speech and answer above the user's reading level. We found no comparable industry programme.
- Research
Neurodivergent people already use chatbots as cognitive prosthetics, and pay for it.
They use them for working memory, task breakdown and emotional regulation. Several studies also report masking, over-reliance, privacy trade-offs under distress and answers that feel too neurotypical.
- Preprint
Models assume more scaffolding, not less load.
Excess length varies by an order of magnitude between assistants. A 2026 preprint found that telling a model the user is neurodivergent makes its answers longer, not lighter.
- Standard
Regulation is arriving without a design brief.
The EU AI Act ties high-risk AI to EU accessibility law (Art. 16(l)) and requires the AI disclosure itself to be accessible (Art. 50(5)). Annex III obligations now apply from 2 December 2027. None of this defines an accessible conversation.
Checklist: accessible AI conversation
Eighteen practices for teams designing chatbots, voice assistants and AI features. Each one shows what it rests on, so you can tell a published standard from our own reading of the evidence.
Lead with the answer, expand on request
Why. Long default answers add reading and working-memory load. Excess length varies widely between models, and autistic users in one study valued quick, digestible answers.
How. Give the direct answer first in a few sentences, then offer optional detail ('More?') instead of front-loading everything.
Research COGA pattern 'Keep Text Succinct'; Jang et al. CHI 2024; YapBench (preprint, 2026)
Offer a simpler-language mode users control
Why. Simplified text improved comprehension in a large study. NAUR asks for a simpler-language mode on request.
How. Provide a persistent toggle or command for plain or easy-read output, and keep it on until the user turns it off.
Standard NAUR REQ 21a/21c; COGA 'Use Clear Words'; WCAG 2.2 SC 3.1.5; Guidroz et al. 2025 (preprint)
Never impose silent time limits
Why. Timeouts are a main cause of errors for adults with intellectual disabilities and cut off people who stutter or pause.
How. Remove response time limits, or make them adjustable and extendable, and warn before any session expires.
Standard NAUR REQ 20a; WCAG 2.2 SC 2.2.1, 2.2.6; Nadeeshani et al. CHI 2026
Tolerate pauses, disfluency and restarts
Why. Voice systems treat pauses as the end of a turn, which truncates people who stutter and people with dementia. Fixing endpointing cut truncation by 79% in one study.
How. Tune end-of-speech detection per user, allow a 'keep listening' option, and repair incomplete utterances instead of forcing the user to start again.
Research Lea et al. CHI 2023; Addlesee & Damonte CUI/Interspeech 2023
Confirm uncertain input before acting
Why. Misrecognition is common for atypical speech, and acting on it can cause errors that are hard to undo.
How. When recognition confidence is low or the action matters, read back the interpreted request and ask for a simple yes or no.
Standard NAUR REQ 13b; WCAG 2.2 SC 3.3.4
Support personalised atypical-speech recognition
Why. Generic speech models perform worse for people with dysarthria, stuttering and other speech differences. Personalised data clearly improves accuracy.
How. Offer on-device speech adaptation or integrate models trained on atypical speech, and let users add custom trigger phrases.
Standard NAUR REQ 12a/12b; Speech Accessibility Project; Apple Listen for Atypical Speech; Project Euphonia
Let users switch voice, text and visuals
Why. Voice-only systems exclude non-verbal users, and text-only interaction limits others. Different people need different modes at different moments.
How. Allow input by voice, typing or tapping, show spoken responses as text, and allow switching mid-conversation without losing context.
Standard NAUR 6.2 and 6.6.6; Nadeeshani et al. CHI 2026; Carik et al. 2024 (preprint)
Show progress: steps done, steps left
Why. Multi-step conversations rely on memory. NAUR and COGA both say processes should not depend on users remembering.
How. In multi-step tasks, show or say 'Step 2 of 4' and offer a running summary the user can recall at any time.
Standard NAUR section 1.3; COGA Objective 6 'Ensure Processes Do Not Rely on Memory'
Make memory visible, editable and optional
Why. Neurodivergent users use chatbots as working memory, but they also report privacy worries and trade privacy away when distressed.
How. Let users see, edit and delete what the assistant remembers, store stated preferences so they are never re-asked, and default to clear consent.
Emerging practice COGA Objective 8; WCAG 2.2 SC 3.3.7; Giri et al. CHI 2026; LaMPost (privacy concerns)
Use literal language; explain tone on request
Why. Idioms, sarcasm and implied meaning can be misread. Autistic users valued on-demand clarification of tone over tone indicators.
How. Default to literal phrasing, and offer a 'What does this mean / what tone is this?' action for any message.
Research COGA 'Use Literal Language'; Haroon & Dogar TwIPS 2024 (preprint, n=8)
Signal uncertainty and offer verification
Why. Users with brain fog, visual impairment or dyslexia may be the least able to spot hallucinations, and invented details erode trust.
How. Mark uncertain claims in plain words, link to sources, and offer a short 'check this' path instead of confident prose.
Research W3C AI Symposium 2023; Glazko et al. ASSETS 2023; LaMPost ASSETS 2022; Friedman et al. 2026 (preprint)
Offer few options, not long lists
Why. Most participants with dyslexia in one study found long lists of suggestions overwhelming. Too many choices add decision load.
How. Present one recommendation or at most a few options, and let users ask for more.
Research LaMPost ASSETS 2022 (12 of 19 participants); COGA Objective 5 'Help Users Focus'
Keep persona, wording and controls consistent
Why. Unpredictable behaviour forces users to relearn the system. WCAG requires consistency for navigation and help, and the same logic extends to generated behaviour.
How. Fix the assistant's name, tone, command words and layout, and announce any change in behaviour or model.
Our interpretation WCAG 2.2 SC 3.2.3, 3.2.4, 3.2.6; COGA Objective 1
Let users pause or stop streamed output
Why. Text that streams in and auto-scrolls is moving content that can distract and overload. Screen-reader users need to know when a response is complete.
How. Provide stop and pause controls, avoid forced auto-scroll, and announce completion through a status message.
Standard WCAG 2.2 SC 2.2.2, 4.1.3 (applying these to streaming is our interpretation)
Always offer a route to a human
Why. Automated systems fail some users every time. Human help is a core COGA pattern and a WCAG consistency requirement.
How. Show a clear, consistently placed way to reach a person, and never hide it behind repeated bot loops.
Standard COGA 'Provide Human Help'; WCAG 2.2 SC 3.2.6; Be My Eyes human fallback model
Remove engagement dark patterns and sycophancy
Why. LLMs show sycophancy, retention hooks and anthropomorphic pressure. These risks compound for users who rely on the AI for regulation or social contact.
How. Do not design for time-on-app. Allow respectful disagreement and 'beneficial friction', and keep emotional validation separate from factual claims.
Research DarkBench ICLR 2025; Lewis et al. CHI 2025; Ma et al. 2026 (preprint); Giri et al. CHI 2026
Don't hard-code a neurodivergent stereotype
Why. When told a user is neurodivergent, models produce longer, more structured output, and users already report answers that feel 'too neurotypical'.
How. Let each person tune length, structure and tone directly instead of switching on a diagnosis-based preset.
Our interpretation Gupta & Buryi NDBench 2026 (preprint); Carik et al. 2024 (preprint); Gadiraju et al. FAccT 2023
Disclose AI accessibly and test with disabled users
Why. The EU AI Act requires the AI disclosure to be accessible. Research rarely includes people with cognitive disabilities, and simplified output often goes unvalidated by its target users.
How. Give the 'you're talking to AI' notice in plain language and in every mode, and co-design and test with people with cognitive and speech disabilities.
Standard EU AI Act Art. 50(5); Seymour et al. CUI 2023; Martínez et al. 2024
The gap
Nobody has a cognitive-accessibility pattern set for AI conversation.
Response length, pacing, memory, predictability, repair after a misunderstanding, and friction that protects rather than hooks. We're mapping the W3C's eight COGA objectives onto conversation, to test them with neurodivergent people. If you research, build or live with this, build it with us.
Who is working on it
No sources match
Try a shorter word, or clear the filters to see everything again.
Standards & guidance
Who writes the rules for natural language interfaces, and how far those rules actually reach into conversation.
Natural Language Interface Accessibility User Requirements (NAUR) (opens in a new tab)
W3C draft note listing accessibility needs for voice assistants, chatbots and other natural language interfaces. Last updated 2022.
Accessibility of machine learning and generative AI (W3C draft) (opens in a new tab)
Early W3C editor's draft on how machine learning and generative AI affect web accessibility for users, authors and testers.
Making Content Usable for People with Cognitive and Learning Disabilities (opens in a new tab)
W3C guidance with eight objectives and design patterns for cognitive accessibility, including memory, focus and clear language.
WCAG 2.2 (opens in a new tab)
The current W3C web accessibility standard. Its rules on timing, consistency and errors apply to chat interfaces.
WCAG 3.0 (Working Draft) (opens in a new tab)
Draft successor to WCAG, still in development, including requirements on cognitive tasks and algorithms.
EN 301 549 (opens in a new tab)
European standard for accessible ICT products and services, used to show compliance with EU accessibility law.
ISO/IEC JTC 1/SC 35 User interfaces (opens in a new tab)
International standards committee for user interfaces, including accessibility, AAC and easy-to-read language.
Regulation
Laws that already apply to AI systems and services. They say what to achieve, not how to design a conversation.
EU AI Act (Regulation 2024/1689) (opens in a new tab)
EU law on AI. High-risk systems must meet EU accessibility rules, and AI disclosures must be accessible.
European Accessibility Act (Directive 2019/882) (opens in a new tab)
EU law requiring accessible products and services such as banking, e-commerce and e-books, applying since June 2025.
ADA Title II web rule: compliance date extension (opens in a new tab)
US rule requiring state and local government web content to meet WCAG 2.1 AA, with deadlines moved to 2027–2028.
Algorithms, AI and Disability Discrimination in Hiring (opens in a new tab)
US Department of Justice guidance: employers using AI hiring tools must not screen out disabled applicants.
UK public sector website and app accessibility requirements (opens in a new tab)
UK rules requiring public sector websites and apps to meet WCAG 2.2 AA and publish accessibility statements.
EHRC: AI and the Public Sector Equality Duty (opens in a new tab)
UK equality regulator's guidance on how public bodies using AI must consider discrimination and equality.
Research & programmes
Where the money and data are going. Most of it is atypical speech; cognitive accessibility gets far less.
Speech Accessibility Project (opens in a new tab)
University of Illinois project collecting speech from people with disabilities to improve speech recognition, backed by five tech companies.
Project Euphonia (opens in a new tab)
Google Research project improving speech recognition for people with atypical speech, sharing open-source tools for personal models.
W3C AI and Accessibility Research Symposium 2023 (opens in a new tab)
W3C event and report on the benefits and harms of AI for people with disabilities.
AI and Accessibility Skills (Southampton and Jisc) (opens in a new tab)
UK research project and white papers on critical AI and accessibility skills, with workshops running to 2027.
Microsoft Accessibility: AI innovation (opens in a new tab)
Microsoft's accessibility hub covering responsible AI guidance and an AI skills course taught by disabled people.
Tools & features
Shipped features and apps that make AI conversation work for more people, plus one benchmark for dark patterns.
Apple Listen for Atypical Speech (opens in a new tab)
iPhone and iPad feature that learns a person's speech patterns on the device to improve recognition.
Alexa Adaptive Listening (opens in a new tab)
Alexa setting that gives people more time to finish speaking before the assistant responds.
Voiceitt (opens in a new tab)
Speech recognition app for non-standard speech, offering dictation, spoken output, captions and a ChatGPT connection.
Be My AI (Be My Eyes) (opens in a new tab)
AI image descriptions for blind and low-vision people, with human volunteers available for harder requests.
Seeing AI (opens in a new tab)
Free Microsoft app for blind and low-vision people that reads text and describes photos and surroundings.
goblin.tools (opens in a new tab)
Free single-task AI tools, such as task breakdown and tone checking, built mainly for neurodivergent people.
DarkBench (opens in a new tab)
Research benchmark that tests chatbots for six dark patterns, including sycophancy, user retention and anthropomorphism.
Common questions
The short answers.
Is there a standard for accessible chatbots or voice assistants?
Not one that covers today's AI assistants. W3C's Natural Language Interface Accessibility User Requirements (2022) is the closest, but it predates large language models. WCAG 2.2 and the COGA guidance still apply to timing, consistency, errors and clear language in chat interfaces.
Does the EU AI Act require accessibility?
Yes, in two places. Providers of high-risk AI systems must meet the accessibility requirements of the European Accessibility Act and the Web Accessibility Directive (Art. 16(l)), and the notice that you are talking to an AI must itself be accessible (Art. 50(5)).
What makes an AI conversation cognitively accessible?
Short answers first with detail on request, no silent time limits, tolerance for pauses and restarts, visible progress in multi-step tasks, memory the user can see and edit, literal language, few options at a time, consistent behaviour, and a route to a human. Our checklist lists 18 practices with the evidence behind each.
Should an assistant change its answers if it knows a user is neurodivergent?
Not by diagnosis. One 2026 preprint found models respond to that label with longer, more structured answers. It is better to let each person set length, structure and tone directly.