Should you allow AI in job interviews? A policy framework for hiring leaders in 2026

There is no universally-correct policy on AI in job interviews. Canva's engineering team tells candidates to use Copilot, Cursor, and Claude during their backend, frontend, and ML coding rounds, and writes that they "insist" on it. Amazon explicitly bans generative AI tools during interviews and calls their use an unfair advantage. Both companies are defensible.

The right policy for your hiring process depends on four variables: the role you are hiring for, the interview stage, the regulatory context, and what signal the round is actually supposed to produce. Picking a side dogmatically (always allow or always ban) is bad management because the answer changes meaningfully across those four variables.

This is the policy-framework version of the AI-in-interviews question. The output is a one-page template you can adapt, not a take.

Key Takeaways

• There is no universal right answer. A blanket "always allow" or "always ban" loses you signal or candidates in some context, every time.

• The right policy depends on four variables: the role, the interview stage, the regulatory context, and the signal the round is supposed to surface.

• Three coherent policies fit different mixes of those variables: always allow (Canva model), always ban (Amazon model), and structured allow with rules (the middle path most companies should actually run).

• The legal layer is real. NYC Local Law 144, Illinois AI Video Interview Act, and the EU AI Act change what you can require, detect, and disclose. Skipping this layer is the most expensive mistake in the policy.

• Treating AI use and AI cheating as the same thing is a category error. The policy must distinguish them clearly or it will lose you good hires.

The reason every "should you allow AI" listicle gets it wrong

Most articles on this topic pick a side. Canva says you must allow it. Amazon says you must ban it.

A Medium essay from a candidate calls the AI penalty real and warns against using AI in coding interviews. A vendor selling a detector argues that AI use is cheating that must be caught.

All four are right inside their own context and wrong as universal policy. A coding role at Canva where engineers use AI all day deserves an interview where AI is allowed openly.

A compliance auditor role at a bank, where the real job involves reading regulatory filings against rules the bank can defend in court, deserves an interview where AI is restricted and the restriction is auditable. A take-home for a security researcher deserves a different policy again.

The honest claim of this article: the right policy is a function of variables you can actually measure, and most companies are writing the wrong policy because they are skipping the variables entirely. The framework below is the four variables and the three coherent policies that fit different mixes of them.

The four variables that shape the right AI interview policy

A real policy is a function of four things. None of them is "AI good" or "AI bad." Each is a question you can answer about your hiring context, and the policy falls out of the combination.

Variable 1: the role

Two questions to answer about the role before the interview policy is written.

Does AI use figure into the actual day-to-day? A backend engineer who will use Copilot daily lives in a different policy world than a financial regulator who will be auditing AI-generated filings under a regulator's eye. The question is not whether AI exists in the role but whether the candidate's first month involves it.

What does responsible AI use look like in that role? A senior PM using AI to draft strategy memos still owns the strategy. A junior support agent using AI to draft customer replies still owns the customer outcome. A pilot using AI flight-control suggestions still owns the aircraft. "Responsible use" is a specific behaviour, role by role: knowing when the model is wrong, owning the output, debugging when it breaks. Your interview policy should make that specific behaviour visible.

If you can't write a one-sentence definition of "responsible AI use in this role," your policy will be confused, because the candidate has no way to know what you're scoring.

Variable 2: the interview stage

Three rounds in a typical loop, three different policy answers.

Screening / phone screen. The signal you're after is basic competence and communication. If the round is a textbook coding question, AI use is trivially game-able, so a "ban without detection" policy is unenforceable. The right move is usually to either redesign the round around an AI-allowed question, or to keep it tight, ban AI use, and accept the signal limitation.

Middle of the loop (coding, system design, behavioural). The signal is judgment, decomposition, trade-off articulation. AI policy here should match the role question above. If AI is part of the day-to-day, allow it openly with rules. If AI is restricted in the role, ban it with a clear disclosure to the candidate up front.

Final round / hiring manager / leadership. The signal is decision-making under pressure, ownership, ability to recover from a hard question. AI use during this round is almost never relevant; the conversation is the test. Default to "no AI assistance, but candidates may take a 30-second pause when they need it" rather than to a detection-heavy policy.

The point of this variable: a single policy across all rounds is the wrong policy. Per-round policy is the right one.

Variable 3: the regulatory context

The legal layer changes the cost and viability of every policy choice. Three frameworks matter for most hiring teams in 2026.

NYC Local Law 144 treats any tool that substantially assists a hiring decision as an Automated Employment Decision Tool. If your "ban AI" policy is enforced via a detection tool, the detector is an AEDT and triggers an annual bias-audit requirement and a candidate notice. The cost is real and the notice changes the candidate experience.

The EU AI Act classifies hiring AI as high-risk. Both AI-conducted interviews and AI cheating-detection systems fall in scope. The compliance burden (conformity assessment, risk management, human oversight, technical documentation) applies whether you are the hiring team or the vendor. If you operate in the EU, your policy must specify the legal basis for the AI use, the human-oversight model, and the candidate-information notice.

The Illinois AI Video Interview Act requires candidate consent and explanation when AI analyses a video interview of an Illinois resident. The act covers AI-conducted interviews and AI evaluation of interview recordings; "we are running AI to score candidates" is a disclosure your policy must make explicitly.

Other US states have their own versions: the California Employment Fairness rules and the broader US state-by-state hiring AI picture all carry compliance weight.

The cost of getting this layer wrong is enforcement risk plus reputational risk. The cost of getting it right is one legal-review cycle at policy-write time. Pay the second cost up front.

Variable 4: the signal you actually need

The most important variable, and the one most policies skip.

For each round in your loop, write down: "the reason this round exists is to find out whether the candidate can _____." If the blank is "write a working function under time pressure," the round is the kind of round candidate-side AI tools were built to defeat.

Your AI policy in that round can be "ban it" or "allow it," but in either case the signal the round is supposed to produce is weak in 2026, and you should consider whether to keep the round at all.

If the blank is "decompose an ambiguous product requirement, choose a sensible architecture, and defend the trade-offs against a senior interviewer's pushback," the round is largely AI-immune at the conversation level. Allow AI as a thinking aid; the signal lives in the conversation.

If the blank is "demonstrate the specific judgment we need on the job, which involves heavy AI use," then banning AI in the round produces no useful signal at all. Allow it openly with rules.

If the blank is "demonstrate the specific judgment we need on the job, which involves restricted AI use," then allowing AI in the round produces a wrong signal. Ban it with disclosure.

The pattern: your policy should be consistent with the signal you actually need from each round, which depends on the work itself. Inconsistency between policy and signal is what produces the worst hiring outcomes.

Three coherent policies, and when each is right

A real AI interview policy fits some mix of the four variables. Three coherent shapes show up across companies that have published their thinking.

Policy 1: always allow, openly. The Canva model. Candidates use AI tools during the relevant rounds; the rounds are redesigned around problems where AI is part of the work. The bar is judgment, ownership, and critical review of generated output. This fits roles where AI use is normal and deep in the day-to-day, and where the company is willing to invest in redesigning rounds. The biggest cost is the redesign work; the biggest benefit is a clean, audit-friendly process with no escalation arms race.

Policy 2: always ban, with disclosure. The Amazon model, more or less. Candidates are told up front that AI use during interviews is prohibited; the rounds are kept tight and live; the prohibition is enforced through conversation pacing and structured follow-ups, not through covert detection. This fits roles where AI use is restricted or audited in the real job and where the company wants the round to produce a "without-AI" baseline signal. The biggest cost is the signal you give up on AI-augmented capability; the biggest benefit is regulatory simplicity and a clean candidate-side experience.

Policy 3: structured allow with rules. The middle path. AI use is allowed in specific rounds, prohibited in others, and the allowed-rounds carry written rules: which tools, what disclosure, how the use is scored, what counts as misuse. This is the policy most companies should actually adopt because most companies hire across a mix of roles and need a per-round answer rather than a single binary. The biggest cost is the writing of the rules and the training of interviewers; the biggest benefit is that the policy matches the variability of the actual hiring.

A clean rule of thumb: if your headcount is in one role family with a uniform AI-use profile, pick policy 1 or 2. If your headcount spans multiple role families with different AI-use profiles, you need policy 3.

The candidate-side problem your policy has to address

Two things are true at once in 2026.

The first is that real-time, candidate-side AI overlays are mainstream. Interview Coder, Cluely, Final Round AI, LockedIn AI, Linkjob, and Parakeet all market themselves as "invisible to screen sharing" and target millions of users between them. The 2023 HireVue survey found more than 85% of HR leaders worried about virtual-interview cheating; the 2026 version of that worry is operationally justified.

The second is that treating "candidate uses AI" and "candidate cheats" as the same thing is a category error that loses you good hires. Many strong engineers use AI assistance as part of their normal workflow. A policy that bans the workflow rather than the deception conflates the two and ends up filtering out candidates whose only "crime" is that the tool the policy banned is the tool their day job runs on.

The policy has to make the distinction explicit. The right framing is not "AI use is allowed" or "AI use is banned"; it is "transparent, role-relevant AI use is allowed under these rules; covert, deceptive AI use that hides the candidate's actual competence is not allowed and will end the interview." The first is a workflow; the second is the actual problem your policy is supposed to address.

Detection of the second kind (covert deception) works inside a well-designed interview through the conversational tells: response latency, eye-movement patterns, voice and confidence step-changes between the prepared answer and the unprepared follow-up. The policy framework above does most of the work to prevent the first kind (transparent, role-relevant use) from being mistaken for the second. Both pieces matter, and they only work together.

A one-page AI-in-interviews policy template

Use this as a starting point. Adapt to your specific roles, jurisdictions, and risk tolerance.

Section 1: roles in scope. List the role families this policy covers. Note which families this policy does not cover (regulated roles often have their own).

Section 2: per-round position. For each round in the loop, state: "AI use is allowed / restricted / banned in this round, because the signal the round is meant to produce is _____ and the candidate's actual job involves _____ AI use."

Section 3: allowed tools and disclosure. If allowed, name the tool categories permitted (coding assistant, web search, model chat). State whether the candidate must disclose use. If banned, state the form of the disclosure to the candidate, the language the recruiter will use, and the consequence of detected misuse.

Section 4: evaluation rubric. State what is scored. Make the rubric public to the candidate at the start of the round, including the criteria that pertain to AI use specifically (e.g., "critical review of generated output," "ownership of the final decision").

Section 5: candidate notice and consent. State the jurisdictions where the policy operates. For each, specify the candidate-facing notice (Illinois AI Video Interview Act consent, NYC LL144 AEDT notice, EU AI Act information notice if applicable). Link to the policy on the careers page.

Section 6: interviewer training. State how interviewers are trained on the policy. Specify the recovery script when a candidate appears to be using AI covertly: pause, restate the rule, give one warning, end the interview if it continues. Train interviewers on the difference between AI use and deception.

Section 7: review cadence. State when the policy is reviewed (quarterly is reasonable in 2026 given the pace of the candidate-side tooling). Specify the trigger conditions for an off-cycle review (a new jurisdiction in scope; a vendor changes capabilities; a regulator publishes new guidance).

A real policy in this format runs 1.5 to 3 pages depending on how many role families it covers. Less than a page is usually too thin; more than 5 is usually too rigid.

Frequently asked questions

Should you allow AI in job interviews? Sometimes. The right answer depends on the role, the interview stage, the regulatory context, and the signal the round is meant to surface. For roles where AI is part of the actual day-to-day work, allowing it openly with rules produces the best signal. For roles where AI use is restricted in the real job, banning it with disclosure is the right call. A blanket policy in either direction usually fits poorly across multiple role families.

Is it ok to use AI in a job interview? Is using AI during interviews cheating? Only when the use is covert and designed to fake competence the candidate does not have. Transparent, rule-following AI use that mirrors the candidate's actual workflow is not cheating; it is a fair test of how the candidate uses tools. The distinction has to be in the policy explicitly or it will be enforced inconsistently.

Can a company ban AI tools in interviews? Yes. Companies are within their rights to set the rules of their hiring process, including banning external tools during the round. The harder question is whether the ban produces a useful signal. If AI use is mainstream in the role's actual day-to-day, the without-AI signal you get from the round may not be a good predictor of on-the-job performance.

Is using ChatGPT in an interview illegal? Generally no, from the candidate side. Some employment contracts or competitive-information rules forbid disclosing confidential information to external tools, which would apply to ChatGPT during certain interviews. On the employer side, deploying AI tools to evaluate candidates is regulated under NYC Local Law 144, the Illinois AI Video Interview Act, and the EU AI Act, among others.

Do you have to tell candidates AI is being used to interview them? In several jurisdictions, yes, with specific notice and consent requirements. Illinois requires consent for AI analysis of video interviews under the AI Video Interview Act. New York City requires a candidate notice when an AEDT is used to substantially assist a hiring decision. The EU AI Act will require an information notice once high-risk hiring AI provisions phase in. For details, the Expert Hire compliance hub is the right starting point.

What happens if a candidate uses AI covertly when the policy bans it? The right response, in the policy, is: pause the interview, restate the rule, give one warning, and end the interview if covert use continues. Do not run a detection theatre; do not publicly accuse the candidate. The point of the policy is to produce a real signal and a fair process, not to win a public confrontation.

What to do about this in your next hiring-process review

The shortest version of the framework: write down the four variables for your top three role families, pick the policy that matches each, and rewrite the candidate notice on your careers page to reflect what you actually do. That is a one-week piece of work for most teams and it materially improves the quality of the signal your hiring process produces in 2026.

If you want the conversational-interview model that fits policy 3 cleanly, look at how the Expert Hire AI interview platform scores candidates and decide whether the rubric structure matches what you would want your interviewers to score against. The same structure works with a human interviewer; the methodology is the point, not the tool.

The right answer to "should you allow AI in job interviews" is not a yes or a no. It is a policy written from the four variables above, applied per round, reviewed quarterly, and grounded in the signal each round is actually supposed to produce. That policy is the one that will hold up against the candidate-side tooling, the legal layer, and the actual hiring you do across the next twelve months.

• By TK, Growth at Expert Hire. Last updated June 11, 2026.