What the proposition assumes
The case for adaptive accessibility is straightforward and appealing. Rather than building one interface and asking every employee to conform to it, deploy systems that adjust themselves to each person, continuously and without being asked. Accessibility becomes personal, immediate, and invisible. The formulation that usually accompanies it is that accessibility should adapt to people rather than requiring people to adapt to it.
The formulation is correct as a principle and misleading as a specification, because it conceals what adaptation requires. A system cannot adjust itself to a person it has not classified. To enlarge text for one employee and not another, to simplify a dashboard, to alter timing, to change input modality, the system must first determine that this particular person needs it. Personalization is classification, and in the workplace the classification at issue is disability.
Personalization is classification, and in the workplace the classification at issue is disability.
That is the governance problem, and it is not a detail of implementation. An employer deploying a system that infers which of its employees are disabled has built something quite different from an accessibility feature. It has built a mechanism that generates disability information about its workforce, continuously, without a request, without disclosure, and without the process the law provides for exactly this purpose.
Classification is the thing employment law most restricts
The Americans with Disabilities Act (ADA) restricts employers from requiring medical examinations or making inquiries as to whether an employee is an individual with a disability, or as to the nature or severity of that disability, unless the inquiry is job-related and consistent with business necessity. The restriction is deliberately broad, and the Equal Employment Opportunity Commission (EEOC) has long read it to cover inquiries likely to elicit information about a disability rather than only direct questions.
Whether an inference engine makes an inquiry within the meaning of the statute is an open question, and it should be labeled as open rather than resolved in either direction. A system that monitors zoom level, dwell time, error rates, input method, correction frequency, or navigation patterns and derives from them that a user is likely to have a visual, motor, or cognitive impairment has not asked anything. It has obtained the answer without the question, which is either outside the provision’s language or precisely the mischief the provision exists to prevent, and no court has yet had occasion to say which.
What is not open is what the employer ends up holding. However the inference is characterized, the output is a record, associated with a named employee, indicating a probable disability. The ADA requires that medical information obtained about employees be maintained on separate forms, in separate files, and treated as a confidential medical record. An adaptive accessibility system generating disability-adjacent inferences inside a general-purpose product analytics pipeline, retained on vendor infrastructure, visible to whoever administers the platform, is unlikely to satisfy that standard, and in most deployments nobody has considered the question because the system was procured as a user experience feature rather than as a source of employee medical information.
Note also what the employee loses. The accommodation process exists to give the person control: whether to disclose, what to disclose, to whom, and what to request. An adaptive system decides all four on the employee’s behalf and never tells her it did. The individual is better served by a plain form and a named decision-maker than by a system that quietly determines what she requires.
The relevant EEOC guidance on these questions, the May 2022 technical assistance on the ADA and the use of software, algorithms, and artificial intelligence to assess job applicants and employees, named impermissible disability-related inquiry as one of three principal mechanisms of algorithmic ADA harm alongside failure to accommodate and screen-out. That document was removed from the agency’s website in January 2025. Its removal changed the availability of the explanation, not the content of the statute.
The market has just been told what happens
Three months before this is written, the Federal Trade Commission announced a complaint and proposed order requiring accessiBe to pay one million dollars to settle allegations that it misrepresented the capability of its AI-powered web accessibility tool. The company had marketed a plug-in claiming it could make any website compliant with the Web Content Accessibility Guidelines (WCAG), in part immediately and in full through an automated process within forty-eight hours, with continuous rescanning to maintain compliance as content changed. The FTC alleged that the product did not do this, failing on basic components including menus, headings, tables, images, and recordings, and that the company had formatted paid endorsements to appear as independent reviews. The proposed order bars the company from claiming its automated products can make a website WCAG-compliant, or keep it compliant over time, absent evidence supporting the claim.
The accessibility overlay is the closest available precedent for adaptive accessibility, and the lesson is not that automation cannot help. It is that automated adjustment layered over an inaccessible underlying system produces the appearance of access rather than access, and that the organizations deploying it acquire the appearance as their defense. Overlays were sold on precisely the register the adaptive accessibility pitch uses now: do not merely comply, innovate; the technology adjusts in real time; accessibility becomes automatic. Hundreds of organizations using overlays have been sued in web accessibility litigation, and the overlay did not help them. It identified them.
An organization that adopts a personalization system on the strength of a vendor’s claim that it adapts to every user is making the same purchase, in a context where the affected people are employees rather than website visitors and where the applicable statute carries an individual right of action.
Adaptation is being offered as a substitute for accommodation, and that is being litigated now
The practical danger is not that adaptive systems fail. It is that their existence is treated as discharging the accommodation obligation.
In March 2025, the American Civil Liberties Union, the ACLU of Colorado, Public Justice, and Eisenberg and Baum filed a complaint with the Colorado Civil Rights Division and the EEOC on behalf of a Deaf and Indigenous employee. According to the public filings, she requested human-generated captioning as a reasonable accommodation for an automated video interview, and the request was denied on the basis that the platform’s automated captions would suffice. Portions of the interview were allegedly uncaptioned in any event. Both companies deny the allegations, and the matter will be determined by investigation rather than by commentary.
Whatever its outcome, the fact pattern names the risk exactly. The employer’s position was that the automated adaptive feature made the requested accommodation unnecessary. That is the proposition adaptive accessibility marketing advances as its central benefit. An organization that adopts the technology and internalizes the marketing has adopted a defense that is currently being tested, and it should understand that it is the defense at issue rather than a way of avoiding the dispute.
Assistive technology is not an accessibility program
A final distinction, because the proposition tends to collapse it. Applications such as Seeing AI and Google Lookout are assistive technologies: tools a blind or low-vision person chooses, configures, controls, and can abandon. They belong to the user. Their governance is the user’s own, and their value is real.
An employer-deployed adaptive interface is a different kind of object. It belongs to the institution, it is applied to the person rather than operated by her, it runs whether or not she wants it, and it generates data the employer holds. Arguing from the success of the first category to the safety of the second is the central analytic error in this field, and it recurs because both are described as AI for accessibility.
There is also a quieter problem in citing consumer assistive apps as evidence of an organization’s accessibility posture. If a blind employee is navigating the workplace using an application she found, installed, and learned on her own time, the organization has not provided accessibility. It has benefited from her doing so, and counting that as institutional provision obscures an unmet obligation. The relevant question is not what tools employees have found. It is whether the organization’s own documents, systems, and physical environment are accessible without them.
What the principle actually requires
Accessibility should adapt to people rather than the reverse. The principle is sound. What follows from it is not inference but self-determination: the person specifies what she needs, and the system does what she specified. That is the difference between a preference control and a detection engine, and it is the whole of the governance question.
Six things follow for any organization considering these systems.
- Ask the vendor, in writing, whether the system infers user characteristics or applies user-declared preferences. If it infers, ask what it infers, from what signals, where the inference is stored, who can see it, and how long it is retained. Retain the answer.
- Prefer declared preference over detection wherever the same outcome is achievable. A settings panel the employee controls delivers the same adaptation with none of the inference problem, and it leaves the employee in charge of her own disclosure.
- Treat any disability-related inference the system generates as confidential medical information for handling purposes, in separate storage with restricted access, and confirm that the vendor’s architecture permits this before deployment rather than after.
- State explicitly, in policy and in the accommodation process itself, that automated adaptive features do not limit or substitute for a reasonable accommodation request, and train whoever handles requests accordingly.
- Refuse compliance claims. Any representation that a product automatically achieves or maintains conformance with an accessibility standard should be treated as an unsupported marketing claim unless accompanied by evidence, which is now also the position of the FTC.
- Fix the underlying system. Adaptive correction applied to an inaccessible document, dashboard, or workflow produces a temporary modification of the surface. Remediating the source is slower, less impressive in a demonstration, and the only thing that durably works.
Conclusion
Personalized accessibility is presented as a way past the friction of the accommodation process: no forms, no disclosure, no delay, no awkward conversation. The friction it proposes to remove is largely the process by which a disabled person exercises control over information about herself and states what she actually needs. Removing it does not eliminate the classification. It relocates the classification from a conversation the employee participates in to a system she cannot see.
The technology is not the problem and much of it will prove genuinely useful. The claim is the problem, and the claim currently being made, that accessibility can be automatic, adaptive, and complete without anyone having to decide anything, is the same claim that has already produced a federal enforcement action and a litigation wave in the adjacent market. It was not true of websites. It is not more true of workplaces, where the people being adapted to are employees with a private right of action and a statute that has protected them since 1990.
References
- American Civil Liberties Union, ACLU of Colorado, Public Justice, and Eisenberg & Baum LLP. Complaint filed with the Colorado Civil Rights Division and the U.S. Equal Employment Opportunity Commission, March 19, 2025.
- Americans with Disabilities Act of 1990, 42 U.S.C. § 12112(d), on medical examinations, disability-related inquiries, and confidentiality of medical information.
- Federal Trade Commission, In the Matter of accessiBe, Inc., File No. 2223156; complaint and proposed order announced January 3, 2025; final order April 2025.
- U.S. Equal Employment Opportunity Commission, The Americans with Disabilities Act and the Use of Software, Algorithms, and Artificial Intelligence to Assess Job Applicants and Employees, EEOC-NVTA-2022-2, May 12, 2022; removed from the agency website January 27, 2025.
- Web Content Accessibility Guidelines, World Wide Web Consortium.