Accessibility and Inclusive AI Design

Recognition failure in a booking assistant is not a scenario to be imagined. It has been measured, on the systems a clinic would actually buy. Koenecke and colleagues ran five commercial recognizers over the same audio: Amazon, Apple, Google, IBM and Microsoft. The audio was 19.8 hours of matched interviews, 42 white speakers and 73 black speakers, recorded across five US sites. The result appeared in PNAS in 2020: "all five ASR systems exhibited substantial racial disparities, with an average word error rate (WER) of 0.35 for black speakers compared with 0.19 for white speakers". Black men were worst affected, at 0.41 against 0.30 for black women.

Now put that recognition layer under a voice assistant that schedules patient appointments. Speech impairments, noisy homes, regional accents, people who need extra time to answer, and a phone line with long waits as the fallback. The benchmark result stops being a benchmark result. It decides who gets an appointment and who gets the queue.

Accommodation after launch handles users one at a time. Accessible-by-design builds the second route into the product. Single-channel optimization quietly picks a winner — and the disordered-speech work shows what the winner's margin is made of.

Speaker-independent models transcribed those 432 speakers at a median WER of 31%. Personalized models reached 4.6% on the same speakers. They also beat expert human transcribers, by a median accuracy gain of 9% and a maximum of 80%. The abstract states it plainly: "The accuracies of personalized models were high (median WER: 4.6%) and better than those of speaker-independent models (median WER: 31%)."

The distance between those two medians is not a limit of the technology. It is a record of whose speech the shipped model was fitted to. Accessible-by-design is a build decision. So is single-channel optimization.

Accessibility asks one question about the people who use a system: can they perceive it, understand it, operate it and recover from it? The people include those with diverse sensory, physical, cognitive, linguistic and technological needs. Inclusive design treats that variation as a core requirement rather than an edge case. Responsible design offers equivalent routes, supports assistive technologies, avoids unnecessary time pressure, preserves user control, and evaluates real tasks with diverse participants.

Accessibility covers the model, the interface, the content, the support process and the remedy. Those layers are not equally governed. WCAG 2.2 and EN 301 549 write testable requirements for the interface and the content. The layer where the recognizers returned 0.35 against 0.19 has no comparable success criterion to fail against. So a team that treats conformance as the whole of accessibility can pass its audit and still lose the patients whose speech the recognizer mishears.

Perceivable, operable, understandable, robust, recoverable. Five properties, and the fifth is where conversational interfaces most often fail. It is also the only one on this map that is already a written requirement rather than a design preference.

That requirement has a number. WCAG 2.2 carries it at Success Criterion 3.3.4, Error Prevention (Legal, Financial, Data), Level AA. It covers pages that cause legal commitments or financial transactions, that modify or delete user-controllable data, or that submit test responses. For those, one of three things must be true. Submissions are reversible — "Reversible: Submissions are reversible". Or entered data is checked, with an opportunity to correct it. Or a mechanism exists for reviewing, confirming and correcting the information before final submission.

Europe makes the same criterion normative. Clause 9.3.3.4 of EN 301 549 says: "Where ICT is a web page, it shall satisfy WCAG 2.1 Success Criterion 3.3.4 Error Prevention (Legal, Financial, Data)." A patient who has booked the wrong slot, cannot reverse it, cannot have the entry checked, and is offered no confirmation step is not meeting a design regret. That is a standards failure with a clause number on it.

Complete the task yourself under each constraint before reading anyone else's accessibility report about it. The clauses below exist so that what you find has somewhere to be filed. They do not substitute for finding it.

Equivalent outcome is the standard, not equivalent effort. Identify the access contexts, define the outcomes each route must deliver, build multimodal control, test with participants, then monitor burden. The later steps are the ones that test whether the fallback costs a user more time, more privacy or more repetition.

Step 4 is the step teams describe as an activity and never size. Here is what it costs and yields at full scale. In 2025 a challenge at Interspeech ran on speech gathered by the Speech Accessibility Project at the University of Illinois Urbana-Champaign: "the 2025 Interspeech Speech Accessibility Project (SAP) Challenge was launched, utilizing over 400 hours of SAP data collected and transcribed from more than 500 individuals with diverse speech disabilities". The top team reached a word error rate of 8.11%, with a semantic score of 88.44%.

The project itself had recordings from about 2,000 participants by the end of September 2025. It is funded by Amazon, Apple, Google, Meta and Microsoft — the same industry whose shipped systems produced the disparity in the case above. Testing with disabled participants at that scale is a dataset and a leaderboard, not a demonstration. A team that cannot reach it should at least know what number it is not measuring against.

Accessibility that survives one release and breaks at the next was never a property of the system, only of that build. The same thing happens to deadlines. At sector scale that is a matter of public record.

On 24 April 2024 the US Department of Justice published its ADA Title II rule on web and mobile accessibility, effective 24 June 2024. It adopted WCAG 2.1 Level AA as the technical standard for state and local government web content and mobile apps. Compliance was due on 24 April 2026 and 26 April 2027. Then, on 20 April 2026, the Department published an interim final rule that pushed both dates back a year: "On April 20, 2026, the Federal Register published the Department's Interim Final Rule (IFR) extending the compliance date for State and local government entities with a total population of 50,000 or more to April 26, 2027." The later date moved to 26 April 2028. A requirement that can be moved a year is not a schedule a team can plan a rebuild around.

Set the completion, fallback and abandonment numbers that force the team to redesign, restrict, remedy or retire the system. Your own thresholds, rechecked after each release — because the external date may not be where you left it.