Assessment design
Do ADA Accommodations Apply to an AI-Based Assessment, and What Do You Offer?
The ADA reaches an AI-based assessment exactly as it reaches a paper test, because both are tests used to make an employment decision: accommodate a known disability unless that is a significant difficulty or expense, and administer the test so a disability does not become what the result measures. Change the conditions (more time, another input or output channel, assistive technology, a typed path), never the skill the assessment exists to measure. Where a request would remove that skill, the answer is a different route to the same evidence.
The takeAn accommodation request is a rubric question wearing legal clothes. It asks what this assessment measures, and teams that can answer that in one sentence settle the request in a morning. The rest find out their exercise was timed because a calendar was, not because speed is the job. Refusals written in the name of the construct are, as often as not, protecting a habit nobody wrote down. So take the request as the free audit it is. An assessment that cannot say what its clock measures was never defensible, whether or not anybody asks for more time.
Where Olive fits
Open a role and see what the work shows
If you are building the assessment yourself, the alternate path is part of the design rather than a form bolted on at the end. Olive's think-aloud takes typed annotation as a first-class equivalent to speech, the assignment pauses and resumes within 48 hours without the paused time counting, and a human reviewer writes all six findings against moments in the session.
Rank your shortlistDoes the ADA apply to an AI-based assessment?
Yes. The ADA covers tests used to make employment decisions, and nothing about the software administering one changes that. The binding text is 29 CFR 1630.11: tests must be selected and administered so results accurately reflect the skill or aptitude the test purports to measure, rather than reflecting an applicant's impaired sensory, manual or speaking skills, except where those skills are the factors the test measures 1.
Two separate obligations sit in that sentence and buyers routinely collapse them into one. The first is administration: the format and manner of the test must not put a disability between the candidate and the thing being measured. The second is accommodation on request. An employer may not fail to make reasonable accommodations, including in the administration of tests, to the known limitations of an otherwise qualified applicant, unless the accommodation would impose an undue hardship 2. A perfectly accessible interface does not discharge the second duty, and granting one request does not discharge the first.
The guidance gap is real and worth naming. The regulation quoted above names no technology at all, and the most direct federal guidance still standing on algorithmic hiring tools (the Justice Department's 2022 document) is written for state and local government employers rather than private ones 3. Private employers can still read its practice list as the clearest published statement of what the duty looks like in an automated process, because the underlying obligation does not change with the vendor.
"AI-based assessment" also covers instruments with almost nothing in common: a model-scored video interview, an adaptive multiple-choice test, a timed coding environment, an AI-open work sample where the candidate uses an assistant. The accommodation surface differs in each. Settle what the assessment is actually measuring first, because every request you receive will be judged against that sentence.
What separates a condition from the construct?
Write down what the assessment measures, in one sentence, before anyone asks for anything. A change that alters how a candidate reaches the task (the input device, the file format, the room, the clock where the clock is not the point) changes the conditions. A change that removes the thing your sentence names alters the construct, and that is the line the regulation itself draws with its exception for skills the test is there to measure 1.
The EEOC puts the same rule in plain words for applicants: employers must give application tests in a format or manner that does not require use of an impaired skill, unless the test is designed to measure that skill 4. Note what that does and does not license. It does not license refusing a request because the assessment was hard to build, or because the alternate format is inconvenient to score. It licenses exactly one refusal: the accommodation would remove the measurement. Even then the obligation continues, as the next section covers.
Time is where this gets decided most often, and where most employers guess wrong in both directions. Extending the clock on a timed coding exercise can change the construct, but only if working under that clock is genuinely part of what the role requires and what the exercise was built to observe. Extending it on a written analysis exercise almost never does, because the time box there is a courtesy to the candidate and a scheduling constraint on the reviewer, not a measurement. Ask which one you have. If nobody can say what the clock measures, it measures nothing, and speed is not the thing worth scoring anyway.
Four more, worked the same way:
- A screen reader on a document-heavy case. The construct is what the candidate does with the source packet: what they open, what they doubt, what they check. Visual reading is the channel. Accommodate.
- A typed path where the assessment asks the candidate to think aloud. If the construct is reasoning, speech is a channel and typing carries it. If the role is telephone-facing and spoken communication is what the exercise observes, speech is the construct.
- Dictation for a long written deliverable. Usually conditions. Not if the job's deliverable is the written document itself and drafting is what you are sampling.
- A request to use an AI assistant. On an AI-open work sample this is already the design and nothing needs granting. On an exercise built to observe unaided output, it removes the construct, so offer a different route to the same evidence rather than that one.
The pattern in all four: the accommodation question resolves into a question about your own rubric. If you cannot answer it, the rubric is the thing that needs work.
What do you offer, concretely?
Offer the change that touches the conditions and leaves the construct standing. In practice that is a short menu, and it is worth writing down before a candidate asks: more time where time is not the point, a different input or output channel, an interface that works with the candidate's own assistive technology, a pause the clock respects, breaks, and a typed equivalent anywhere the assessment listens.
| What the assessment demands | What to offer | The question to settle first |
|---|---|---|
| Reading a source packet | Screen-reader-compatible files, keyboard-only navigation, text-to-speech | Is reading the skill, or what the candidate does with what they read? |
| Speaking while working | A typed path of equal standing, scored the same way | Is the reasoning the construct, or the speech? |
| A fixed clock | More time on a stated multiplier, a pause that stops the clock | Does the clock measure anything, or does it just bound the ask? |
| Typing a long deliverable | Dictation, speech-to-text, a scribe, an outline-plus-discussion format | Is written output the deliverable the job produces? |
| One unbroken sitting | Scheduled breaks, resume within a stated window | Does continuity carry any part of the measurement? |
| A video or camera step | A version without it | Was the camera measuring anything, or was it there for comfort? |
Then publish the route. The Justice Department's practice list is the most concrete published version: tell applicants what technology is being used and how they will be evaluated, give them enough information to decide whether to ask for an accommodation, and provide clear procedures for requesting one, making sure that asking does not hurt their chance of getting the job 3. Its worked example is an online interview program that a blind applicant's screen reader cannot operate, where an accessible version is the accommodation unless it is an undue hardship 3.
The timing rule depends on that disclosure. An applicant is told to raise a need as soon as they realize they have one 4, and they cannot realize it from an invite that says "a 60-minute exercise." Say what the assessment involves (what is timed, what is spoken, what is captured, what software is needed) in the same message that carries the link, with the request route beside it. An accommodation offered after the session is a retake nobody planned for.
How do you handle a request that would change the construct?
Not by refusing and stopping. The duty is to find an accommodation that works, not to grant the exact one named: if the request would remove what the assessment measures, propose an alternative that preserves it: a different channel to the same task, a different task at the same construct, or the same evidence gathered a different way. Undue hardship means significant difficulty or expense 5, a higher bar than inconvenience.
Documentation has a ceiling and it is lower than most intake forms assume. Where the disability and the need for accommodation are not obvious, an employer may ask for reasonable documentation explaining the disability and why an accommodation is needed 4. That is the whole entitlement: the connection between the limitation and the request. A diagnosis, a treatment history and a physician's full file are not the price of a longer clock, and collecting them creates a record you then have to defend.
Keep the accommodation out of the assessment record. The person who reviews the work needs to know that the conditions differed, so the session is read correctly. A paused clock is not a slow candidate. Nobody in that chain needs the underlying reason, and the request should not travel to whoever makes the hiring call as a fact about the person. Write the separation down, because the default in most tools is that everything lands in one file.
Score the accommodated session against the same rubric rows as every other one. This is where an unwritten rubric fails loudly: if two reviewers cannot agree on an ordinary session, they will certainly disagree on one where the format changed, and the disagreement will be attributed to the accommodation. Reviewer agreement is a property of the rubric, and it has to exist before the alternate format does.
Build the alternate path with the assessment, not after it
An accommodation designed after launch is a workaround, and it shows. The alternate format ends up measuring something slightly different, nobody documents what changed, and the first candidate who uses it gets a result that cannot be compared with anyone else's. Build the second path while you build the first: the typed equivalent of anything spoken, the keyboard equivalent of anything dragged, the untimed variant of anything timed.
That sequencing also settles the question of whether an accommodation is a favor. When both paths exist from the start and both are scored against the same rows, the alternate is a supported way to take the assessment rather than an exception someone has to justify. It costs a design conversation up front and a rewrite of one rubric anchor. It costs considerably more after a request arrives with a deadline attached.
If a vendor is running the assessment, the questions are specific and the answers should be in writing before you sign. Which parts are timed, and what does the timing measure? Does the candidate interface work with a screen reader and with a keyboard alone? Is there a typed path for anything spoken, and is it scored identically? What happens to a candidate who declines a capture step, and does the session continue? Does the reviewer see that an accommodation was in place? And who owns the request route, you or them? Evaluating the vendor on those six answers tells you more about accessibility than any conformance badge.
The instrument you pick sets the size of this whole problem. A structured interview carries a small accommodation surface and a weak claim on what someone can actually do; a work sample carries a large surface and a strong one. That trade-off is worth making deliberately rather than discovering it at the first request, and the comparison between the two is the same argument seen from the other side.
Common questions
Can you deny an accommodation that would change what the assessment measures?
You can decline that specific change, but the conversation continues. The obligation is to provide an effective accommodation, not the one named in the request, so the move is to propose an alternative that preserves the measurement: a different channel to the same task, or the same evidence gathered another way. Refusing outright, with no alternative offered and nothing written down about why, is the version that fails. Record the construct the request would have removed, the alternatives offered, and what the candidate accepted.
How much extra time should you offer on an AI-based assessment?
Decide by what the clock measures, then publish the number. If timing is not part of the construct, a stated multiplier (commonly time and a half or double) is simpler to administer and defend than case-by-case negotiation. If timing is part of the construct, say so in the disclosure and offer breaks or a pause that stops the clock instead. Either way, name the base duration and the extended one before the assessment opens, so a candidate can tell whether the standard version already works for them.
Can you ask for medical documentation before accommodating an assessment?
Only where the disability and the need are not obvious, and only for reasonable documentation explaining the disability and why the accommodation is needed. That is the ceiling. A diagnosis, records, or a full physician's report exceed what the request supports, and holding them creates an obligation to protect them. Route documentation to whoever handles accommodations, not to the recruiter or the hiring manager, and keep it out of the file the reviewer reads.
Does the reviewer need to know that a candidate was accommodated?
The reviewer needs to know the conditions differed, and nothing more. A session that ran on a doubled clock or a typed path reads differently from a standard one, and a reviewer who does not know that will misread pacing as hesitation. Pass along the change to the format, never the reason for it. The person making the hiring decision does not need either, and giving it to them creates the exposure the separation exists to prevent.
What if the assessment already lets candidates use an AI assistant?
Then the most common accommodation request is already the design, and the remaining ones are ordinary: format, input, output, time, pacing. An AI-open assessment does raise its own question: whether a candidate's assistive technology and the assistant conflict, and whether the interface running both is operable with a keyboard and a screen reader. Test that with real assistive technology before launch rather than reading a vendor's conformance claim, because the combination is where these interfaces usually break.
References
- 1. 29 CFR 1630.11 - Administration of tests ✓ ecfr.gov Tests must be selected and administered so results accurately reflect the skill or aptitude the test purports to measure rather than an applicant's impaired sensory, manual or speaking skills, except where those skills are the factors the test purports to measure.
- 2. Employment Tests and Selection Procedures ✓ eeoc.gov Title I forbids failing to make reasonable accommodations, including in the administration of tests, to the known limitations of an otherwise qualified applicant unless the accommodation would impose an undue hardship.
- 3. Algorithms, Artificial Intelligence, and Disability Discrimination in Hiring ✓ ada.gov Written for state and local government employers: accommodations are required during hiring absent undue hardship; employers should tell applicants what technology is used and how they will be evaluated, give enough information to decide whether to request an accommodation, and provide clear request procedures; the worked example is an online interview program a blind applicant's screen reader cannot operate.
- 4. Job Applicants and the Americans with Disabilities Act ✓ eeoc.gov Application tests must be given in a format or manner that does not require use of an impaired skill unless the test is designed to measure that skill; applicants should request an accommodation as soon as they realize they need one; where the disability and need are not obvious an employer may ask for reasonable documentation explaining the disability and why an accommodation is needed.
- 5. 42 U.S.C. 12111 - Definitions (ADA Title I) ✓ law.cornell.edu Paragraph (10)(A) defines undue hardship as an action requiring significant difficulty or expense, judged against the factors in subparagraph (B).
5 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.