The next discriminatory interface may not look discriminatory at all.
It may greet a deaf job applicant with a pleasant video interviewer, measure “communication” from a voice response and quietly rank the applicant below people it can hear. Or it may be a benefits chatbot that reads a question but cannot conduct a conversation in signed language. Federal civil-rights agencies warn that algorithmic tools can screen out people with disabilities without proper safeguards.
America needs a credible AI access label: a voluntary signal showing whether a system has been tested with deaf and hard-of-hearing people, where it fails and who must fix it. Not a membership badge. Evidence, not promises.
This is not pre-clearance. It is an honest market signal. A useful certification creates a separating equilibrium: Firms that do the work can distinguish themselves from firms that merely learn the language of inclusion. The signal must be auditable. Otherwise it is cheap talk — no answer to the only question that matters: can a deaf person actually use this system to get the job done?
Federal procurement already understands the difference. An Accessibility Conformance Report explains how an information and communications technology product conforms to accessibility standards and helps public buyers assess products. Government guidance seeks the evaluation methods behind a claim, inaccessible core functions, user impact and hands-on testing of the actual release.
That is the lesson for AI — and the failure mode to avoid. A conventional report can record contrast and button labels; it cannot establish that a model understands a signer or gives a person a fair chance when it is wrong. Without cohort-specific testing: attest, procure, discover the problem later.
An access label should make that ritual harder. Its requirements should be plain.
It should test signed-language pathways, not treat them as an afterthought to speech recognition. It should test visual and haptic equivalence: a warning or handoff must have comparable timing and meaning, not arrive as a delayed text substitute. NIST notes that people derive meaning from AI output differently and that human-AI interactions vary.
It should publish results disaggregated by user cohort and task. An average score can conceal a severe failure for signing users, users with different language histories or people operating under time pressure. NIST cautions that turning complex human practices into measurable quantities can remove context needed to understand real-world impact.
It should document error modes in high-stakes systems. Which errors are annoying? Which can shut someone out of a job, medical care, education or a public benefit? How often do they occur in the relevant cohort? What does the product do next? The E.E.O.C. and Justice Department say employers need an accommodation process when using algorithmic decision tools; without safeguards, qualified workers with disabilities can be screened out.
And it should require a real human fallback. Not a support email answered next week: a route to a person with authority to review an adverse result, in time for the decision at stake. NIST says some systems require human oversight, roles must be defined and organizations should track when humans overrule AI output and why.
The rest is governance, documentation and privacy. The label should identify who owns accessibility performance after launch; preserve auditable protocols, findings and remediation; and impose strict rules for signer-video data. Consent, retention limits, access controls and a ban on secondary use should be conditions of evaluation, not boilerplate after it.
The label must also be governed in public. Publish the standard, test methods, conflicts and rationale for revisions. Let civil-rights groups, disability technologists, procurement officials, researchers and companies comment in a public docket. Deaf experts need a durable role in design review — not a ceremonial seat after requirements are settled. A label blessed only by a gilded circle of large vendors will be read, correctly, as capture. NIST calls for interdisciplinary, demographically diverse teams and feedback from affected communities.
The Institute for Inclusive AI is well placed to build the label because it joins AI-policy expertise with Deaf-led work on accessible computing, captioning and sign language, and works across research, policymaking and industry. That bridge matters: engineers alone can miss the mechanics of communication; advocates alone may not reach procurement; vendors alone will be captured.
But being well placed is not a blank check. The institute’s materials describe membership-based label support and advisory positions for founding partners. Resolve that tension in the design: membership cannot buy a pass, advisory access cannot set the test and conflicts must be public. If the institute opens its methods, failures, governance and appeals to Deaf users and independent challenge, it can offer what the market lacks: a label that means access has been shown, not simply promised.
%402x.svg){kind=icon}
