Policy

Can You Legally Ask Candidates How They Use AI?

No US law bans asking how a candidate works with AI, and a job-related question about method clears the EEOC's pre-employment rule. Ask every candidate the same question; an inquiry that screens out a protected group may be used as evidence of intent to discriminate, absent a business purpose. Software that scores the answer adds notice, bias-audit and disclosure duties in New York City and Illinois. Asking why someone needs the tool, rather than how, can become a pre-offer disability inquiry the ADA bars.

The takeMy read is that employers have this backwards. The question feels risky and gets dropped, while the tool arrives pre-audited by somebody else and keeps running. A question asked of every candidate in the same words leaves a record you can hand to a regulator. A vendor score leaves you attesting to an audit of a model you cannot open. The AI statutes here were drafted at the tool, and enforcement is most likely to follow the drafting. What the interview carries is one line you cannot cross, the ADA's, and it is the part getting quietly retired.

Where Olive fits

Open a role and see what the work shows

Olive produces no automated decision for a notice to describe: a person writes all six findings, each anchored to a timestamped excerpt from the session, and every released report exports with its rubric, scorer and bank versions attached. The candidate is granted the identical document.

Rank your shortlist

Is Asking About AI Use a Lawful Interview Question?

Yes. No federal, state or city law prohibits asking a candidate how they work with AI, and the EEOC's general rule for pre-employment inquiries is that what you request should be limited to what is essential for determining whether a person is qualified for the job 1. A question about method, on work you are actually hiring for, clears that bar. What it has to avoid is collecting anything else on the way.

The EEOC's own caution names the failure mode. Pre-employment inquiries that relate to or disproportionately screen out members of a protected group may be used as evidence of an employer's intent to discriminate unless the questions can be justified by some business purpose 1. The wording therefore carries weight that the topic does not. "Which AI tools do you use?" gathers a list of subscriptions. "Walk me through the last time an assistant handed you something confident and wrong" gathers work.

Two habits carry most of the protection. Ask every candidate in the requisition the same question at the same stage, because a question put only to the candidates who look like they might not use AI is a question about something other than AI. And record the answer as what the candidate described, in the same place as the rest of your interview notes, rather than as an impression of the person.

The risk is not in the topic. It is in the file: who was asked, what was written down, and what the writing was used to do.

When Does the Answer Become an Automated Employment Decision?

The moment software turns it into a simplified output that carries the call. New York City's rules define substantially assist or replace discretionary decision making three ways: relying solely on a simplified output such as a score, tag, classification or ranking; weighting that output more heavily than any other criterion; or using it to overrule conclusions drawn from other factors, including human decision-making 2. An interviewer weighing a transcript against four other things is none of those.

Cross that line and two duties attach, on two different triggers. The bias audit, no more than a year old with a summary of results published, attaches to an employer or employment agency using the tool in the city 2. The notice, at least 10 business days before the tool is used, runs to a candidate for employment who resides in the city; it can sit on the employment section of your website, in the job posting, or go by mail or email, and the rules accept any of the three 2. The audit is the expensive half, and it is an exercise on your own applicant flow rather than a document the vendor hands you.

Illinois has two rules and they catch different things. The Artificial Intelligence Video Interview Act applies when you ask applicants to record video interviews and run an AI analysis of them: notify the applicant before the interview, explain how the artificial intelligence works and what general types of characteristics it uses to evaluate applicants, and obtain consent. An applicant who has not consented may not be evaluated by it at all 3. Since January 1, 2026 the Illinois Human Rights Act adds a broader duty. Using AI in recruitment or hiring in a way that has the effect of subjecting people to discrimination on a protected basis is a civil rights violation, using zip codes as a proxy for a protected class is a civil rights violation, and so is failing to notify a person that AI is being used for those purposes 4.

Colorado's SB 24-205 took the same substantial-factor shape and added a risk-management program, an impact assessment, notice to the person, a chance to correct inaccurate personal data, and an appeal with human review where technically feasible 5. None of it ever bound anyone. Its start moved from February 1 to June 30, 2026 6, and Governor Polis signed a full repeal and reenactment that May 9. The replacement takes effect January 1, 2027 and drops the risk-management program and the impact assessment, so a compliance summary telling you Colorado binds employers this year is out of date.

None of this reaches an interviewer asking a question and writing down the answer. All of it reaches a transcript fed to a model that returns a fit number.

What Turns a Fair Question Into an ADA Problem?

Asking why the candidate needs the assistant, rather than how they use it. The ADA regulations make it unlawful to conduct a medical examination of an applicant, or to ask whether an applicant is an individual with a disability or about the nature or severity of that disability, before an offer 7. The permitted form is inquiry into the ability to perform job-related functions, including asking an applicant to describe or demonstrate how they would perform them 8.

That line matters more with AI than with most tools, because the same software is assistive technology for a lot of people. Speech-to-text, reading and summarizing support, drafting help for dyslexia: a candidate answering "how do you use AI at work" may be one follow-up away from telling you something you were not permitted to request. Push on why, and you have requested it.

So ask about the task. "Here is a brief like the ones this role gets. Describe how you would take it from here" sits close to the words the regulation itself uses, which permit asking an applicant to describe or demonstrate how, with or without reasonable accommodation, they will perform job-related functions 8. It also produces a better answer than a tour of somebody's toolbar.

If a candidate volunteers that AI is an accommodation, the conversation has changed job. It is an accommodation conversation now, handled by whoever handles those, and it does not belong in the evaluation notes. Running that during a timed exercise has its own mechanics, which is where accommodations on an AI-based assessment comes in.

Write the Question So the Answer Is Evidence

Ask every candidate the same question about work they actually did, and write the scoring anchors before you hear the first answer. Comparability is what makes an answer usable and a decision defensible: one prompt, one set of criteria, notes on what each person said rather than how they came across. The legal shape and the useful shape are the same here.

Three questions do most of the work, and all three are about a task rather than a habit:

  • The first move. "Take the last piece of work you did with an assistant. What did you ask for first, and why that first?" The answer separates people who framed the problem from people who asked for the deliverable.
  • The wrong answer. "When did it give you something confident and wrong, and how did you find out?" The useful half is the finding out: a page opened, a number recomputed, a colleague asked.
  • The boundary. "What did you keep and do yourself on that task, and why not hand it over?" A candidate who kept nothing and a candidate who kept everything are both telling you something.

Leave out how many hours a week they use it, which subscription they pay for, what they use at home, and whether they could do the job without it. None of those is job-related, the first three collect data you have no use for, and the last one walks straight into the ADA problem above. The same test sorts the rest of the guide, and which interview questions are actually illegal, and what replaces them works from the substitute the job needs rather than from a banned list.

Then have a person score it against written anchors. Running the round as a structured interview is the difference between comparable answers and four managers with four impressions, and if you want a longer bank to draw from, twelve questions that show how a candidate uses AI are already written down.

Which Jurisdiction's Rules Apply to a Remote Candidate?

Mostly the candidate's, and in New York City, both. Its notice duty runs to a candidate for employment who resides in the city, while its bias-audit duty follows your own use of the tool in the city 2. Illinois's video-interview statute applies when you consider applicants for positions based in Illinois 3, and the Illinois Human Rights Act reaches any individual performing services for remuneration within that state 4. One remote requisition can sit under two or three sets of duties at once.

The practical answer is not a per-state question matrix. Write one question set that clears the strictest rule you touch, publish one notice, and apply both everywhere. A notice costs a paragraph on the employment section of a careers site, posted ten business days ahead of use 2. A fifty-state decision tree costs a recruiter on every requisition, forever.

Retention travels with the question. On an applicant's request, Illinois requires video interviews to be deleted within 30 days, including all electronically generated backup copies, and requires anyone who received a copy to be instructed to delete it too 3. That is a build item rather than a policy sentence, because somebody has to be able to find every copy.

The honest limit: all of this settles whether you may ask and what you must disclose. It settles nothing about whether the answer predicts anything on the job, which is a separate question with its own evidence. Write down which of the two you are claiming, because the policy you publish is where that sentence belongs.

Read the evidence

Common questions

Can you ask which AI tools a candidate used on a take-home?

Yes, and asking beats guessing. A tool list on its own tells you little, so pair it with the decisions: what the assistant produced, what was changed, and what was checked against something outside the chat. Ask it in identical words of every candidate, and file the answer with the rest of the submission review. Treating an admission as disqualifying while undisclosed use passes is hard to defend: it rewards silence, and it produces a record that is hard to explain to anyone who asks about it later.

Does asking about AI use trigger a bias audit requirement?

Not by itself. New York City's obligations attach to an automated employment decision tool, meaning machine-learning output that substantially assists or replaces discretionary decision making: relying solely on a simplified output, weighting it above every other criterion, or using it to overrule human conclusions 2. An interviewer asking a question and writing notes sits outside that definition. Feed those notes to a model that returns a fit score and both duties attach at once, the bias audit and the ten-business-day notice.

What if a candidate says they use AI because of a disability?

Stop evaluating and start accommodating. Before an offer, an employer may not ask whether an applicant has a disability or about its nature or severity 7. A volunteered disclosure is not a violation by you, but acting on it is. Hand the accommodation request to whoever handles those, keep it out of the interview notes and out of the hiring file, and carry on assessing the work. What the candidate produced, and how they checked it, are still the things being judged.

Is a written AI-use disclosure on the application legal?

Generally yes, on two conditions. Ask it of every applicant for the requisition, and be specific about scope, because a question about this application is answerable and a question about somebody's AI use in general is not. Say what the answer will be used for in the same breath: a disclosure question with no stated purpose reads as a trap and gets answered like one. If the answer feeds an automated screen rather than a human reader, the automated-decision rules in your candidates' jurisdictions apply 24.

Can you require disclosure and then reject someone for using AI?

Yes, though it is worth deciding why before you do. Rejecting for disclosed use while identical undisclosed use passes selects for candidates who did not answer honestly. If AI is genuinely barred for a role, because of a client contract or a handling restriction, say so in the posting and before the work rather than at the decision. Whether to allow it at all is the question underneath this one, and the posting is where it gets answered.

Should the AI question go in the screen or in the interview?

The interview, where a follow-up is possible. In a screen it collects a claim, and a resume line reading AI-proficient and an application checkbox are the same artifact: neither survives one specific question about a real task. Put it where an interviewer can ask what the assistant got wrong and how the candidate found out. If volume makes that impossible, use a short work sample instead of a longer questionnaire, because an answer about method is worth less than a record of the method.

References

  1. 1. Prohibited Employment Policies/Practices U.S. Equal Employment Opportunity Commission, 2026. eeoc.gov The general rule that pre-employment information requested should be limited to what is essential for determining whether a person is qualified for the job, the explicit prohibition on pre-offer disability inquiries, and the caution that inquiries which relate to or disproportionately screen out a protected group may be used as evidence of intent to discriminate absent a business purpose.
  2. 2. Notice of Adoption of Final Rule: Automated Employment Decision Tools (6 RCNY 5-300 to 5-304) NYC Department of Consumer and Worker Protection, 2023. rules.cityofnewyork.us The three-part definition of substantially assist or replace discretionary decision making and the definition of a candidate for employment (5-300); the annual bias audit and published-results duties on an employer or employment agency using an AEDT in the city (5-301(a), 5-303(a)); and the three permitted ways of giving notice at least 10 business days before use to a candidate who resides in the city (5-304(b)).
  3. 3. Artificial Intelligence Video Interview Act (820 ILCS 42/) Illinois General Assembly, 2019. ilga.gov Section 5 requires notice, an explanation of how the artificial intelligence works and what general types of characteristics it evaluates, and consent before an AI analysis of a video interview for a position based in Illinois; Section 15 requires deletion within 30 days of an applicant's request, including backup copies.
  4. 4. Illinois Human Rights Act, Article 2 - Employment (775 ILCS 5/2-101, 5/2-102) Illinois General Assembly, 2026. ilga.gov Subdivision 2-102(L), added by P.A. 103-804 effective 1-1-26, makes it a civil rights violation to use AI in recruitment or hiring with the effect of discriminating on a protected basis, to use zip codes as a proxy, or to fail to give notice of that use; 2-101(A) defines an employee as an individual performing services for remuneration within the state.
  5. 5. SB24-205: Consumer Protections for Artificial Intelligence Colorado General Assembly, 2024. leg.colorado.gov Deployer duties for a high-risk AI system used in a consequential employment decision: a risk management policy and program, an impact assessment, notice to the consumer, an opportunity to correct inaccurate personal data, and an appeal with human review where technically feasible.
  6. 6. SB25B-004: Increase Transparency for Algorithmic Systems Colorado General Assembly, 2025. leg.colorado.gov Extends the effective date of the requirements of Senate Bill 24-205 to June 30, 2026.
  7. 7. 29 CFR 1630.13 - Prohibited medical examinations and inquiries EEOC ADA regulations, eCFR, 1991. ecfr.gov Except as permitted by section 1630.14, it is unlawful to conduct a medical examination of an applicant or to make inquiries as to whether an applicant is an individual with a disability or as to the nature or severity of such disability.
  8. 8. 29 CFR 1630.14 - Medical examinations and inquiries specifically permitted EEOC ADA regulations, eCFR, 1991. ecfr.gov Paragraph (a): a covered entity may make pre-employment inquiries into the ability of an applicant to perform job-related functions, and may ask an applicant to describe or demonstrate how, with or without reasonable accommodation, the applicant will be able to perform job-related functions.
  9. 9. SB26-189 Automated Decision-Making Technology - Bill Summary (as enacted) Colorado General Assembly, 2026. leg.colorado.gov The repeal and reenactment of SB 24-205 before it applied, and the January 1, 2027 trigger for the replacement.

9 sources, numbered by first appearance. How Olive sources claims

General guidance, not legal advice. Hiring rules differ by state and country and change often; check anything here against your own counsel before you act on it.

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.

Back to answers

Open your first role Ten attempts a month against a live item bank, with a human-written report on every one.