TL;DR: With candidates using AI in interviews, recruiters mostly cannot tell: in the one controlled test this page found, no interviewer suspected AI help. AI-text detectors misclassify non-native English writing, so the practical decision is a policy, not a detection tool. Employers have three real options: prohibit live AI use, allow it with disclosure, or redesign the interview around AI. This page gives a matrix and sample wording for each.
Updated 28 October 2026; policy examples checked against Anthropic's guidance (last revised 10 Jul 2025).
Key takeaways
- Gartner's 4Q24 survey of 3,290 candidates found 39% used AI during the application process. That figure covers applications, not live interviews (Gartner, 31 Jul 2025).
- 59% of 3,000 US hiring managers in Checkr's September 2025 survey suspected a candidate of using AI to misrepresent themselves. "Suspected" is not confirmed, and the survey is a vendor's.
- In interviewing.io's test (37 interviews, 32 analysed), no interviewer suspected AI help, and the cheater pass rate fell from 73% on verbatim questions to 25% on custom ones.
- Real-time assistants advertise invisibility to screen share. Those are sellers' marketing claims, not test results.
- The defensible choice is a published policy matched to what you are measuring, with the interview structured so AI help matters less.
Gartner's survey found that 39% of candidates used AI somewhere in the application process. In the one controlled test of live interviews this page found, no interviewer noticed AI help at all. Neither measures how often candidates use AI in live interviews, but together they frame the question this page addresses: what you can and cannot tell, and what rule to set.
This page is about candidates using AI to help with answers. Impersonation and identity fraud belong to candidate fraud in remote hiring, and the two problems need different responses.
How common is candidate AI use, and what does the data actually cover?
Three sources give figures. None of them measures AI use in live interviews directly.
- Gartner, 4Q24 survey, n=3,290. 4 in 10 candidates used AI during the application process (Gartner press release, 31 Jul 2025; HR Dive, 8 Aug 2025). This is the application stage. Keep it separate from live interviews.
- Fabric, July 2025 to January 2026. According to Fabric, across 19,368 AI-led interviews on its own platform, 38.5% of candidates were "flagged for cheating behavior" (Fabric). This is vendor-reported, and flagged is not confirmed.
- Checkr "Hiring Hoax," 16 Sep 2025, n=3,000 US hiring managers. 59% suspected a candidate of using AI to misrepresent themselves, and 62% agree that job seekers are now better at faking their identities with AI than hiring teams are at detecting them (Checkr). It is a vendor survey, and "suspected" is not "confirmed."
Can recruiters tell if a candidate is using ChatGPT?
Usually no. The only controlled test this page found had interviewers evaluate candidates who were using ChatGPT, and in 32 analysed interviews none of them suspected cheating.
The detail, from interviewing.io: 37 interviews were run and 32 analysed. The pass rate was 73% on verbatim LeetCode questions, 67% on modified ones and 25% on custom questions, against 53% for the control group. The experiment is undated (about 2023) and the sample is tiny, so it shows a direction, not a rate.
Vendors list signs that an interviewer might observe: response latency patterns, reading-like gaze, unnatural cadence, and answers that collapse under an unscripted follow-up (Glider, 21 Aug 2026; Alex, 8 Oct 2025; Fabric). All three are vendor-authored, and none of the pages we read (15 Oct 2026) publishes accuracy figures.
What is not reliably observable are overlays and audio-capture assistants. Sellers claim invisibility to the interviewer: one homepage reads "Invisible to screen share," and another claims its tool is "100% invisible on Zoom, Google Meet, Teams, HackerRank, CoderPad" (Cluely and Final Round AI homepages, fetched 15 Oct 2026). These are marketing claims, not test results, and we have not independently verified them. The claim-against-evidence detail is in real-time interview coaching tools. This page does not recommend those tools and does not describe how to evade review.
Why AI-text detectors are the wrong tool
WeAreDevelopers states that tools for detecting AI text "don't work any more" (WeAreDevelopers). That is an undated opinion from a job board, not a test, so we treat it as a report.
Detectors also raise a bias concern. Research (Liang et al.) found that detectors misclassify writing by non-native English speakers (arXiv 2304.02819). The study tested essays, not interview speech, so it does not transfer directly.
Do not use detectors as a control. If the concern is a person who is not who they say they are, that is a separate problem, and the fake-candidate red flags cover the identity side.
Your three policy options (decision matrix)
Preparing with AI beforehand is normal, and it is neither detectable nor objectionable. Real-time AI during an evaluation of present ability is the contested line. Employers can rationally choose any of three responses, and the choice depends on what the interview is meant to measure. The position taken here is that the choice should be made on purpose, written down, and told to candidates. Where candidate AI use fits into the wider employer-side decision is covered in AI in the hiring process.
| Prohibit live AI | Allow and disclose | Structure the interview around AI | |
|---|---|---|---|
| What you are measuring | Unaided ability in the live conversation | How well the candidate works with AI and explains the result | Skill that holds up when AI is available |
| How you would enforce | State the rule in the invitation; use unscripted follow-ups; human review of any flag | Ask candidates to say which tools they use and to show their work | Design tasks AI cannot shortcut: out-of-distribution problems, building one's own tools, live follow-ups |
| Detection burden | High; overlays and audio assistants are not reliably visible | Low; the rule is disclosure, not detection | Low to moderate; the structure carries the load |
| Candidate-fairness risk | Flags can land on gaze, latency and pacing differences linked to disability or neurodivergence | Uneven access to tools | Tasks must be accessible; offer an accommodation route |
| Best for | Roles where unaided live reasoning is the job | Senior and technical roles where AI is part of the work | Technical roles and any stage already exposed to AI help |
| Sample policy sentence | "Please do not use AI tools or other assistance in live interviews unless we say so." | "You may use AI tools if you tell us which ones at the start." | "We will ask you to explain and extend your answers, so the interview rewards your own reasoning." |
| Real example | Anthropic's live-interview rule; Amazon, as reported by Fortune | No employer example is sourced on this page. | Anthropic's take-home redesign |
The "Best for" row is editorial judgment, not drawn from a source. The examples are explained in the next two sections.
What have employers actually published?
Anthropic's candidate guidance, last updated 10 Jul 2025, sets rules by stage (Anthropic):
- Applications: draft yourself, then refine with Claude.
- Take-home exercises: no Claude unless the exercise says so.
- Interview preparation: encouraged.
- Live interviews: "This is all you–no AI assistance unless we indicate otherwise."
Fortune reported on 11 Jun 2025 that an Amazon recruiter directive said candidates found using AI may be disqualified from the process (Fortune). We rely on that report and have not seen the directive itself.
How do you structure the interview around AI?
Anthropic Engineering's Tristan Hume described redesigning a take-home test several times as Claude models matched top human scores. The lessons were to use "out of distribution" tasks and to require candidates to build their own tools, and he added that "Realism may be a luxury we no longer have" (Anthropic Engineering, 21 Jan 2026).
The same pattern shows up in the interviewing.io results: custom questions cut the cheater pass rate from 73% to 25%. In practice, that means:
- Unscripted follow-ups on the candidate's own answer.
- A short live-reasoning segment, where the candidate works through a problem aloud.
- An in-person round for finalists, where it is proportionate to the role.
Detection software has a place, with limits. It can log events and keep records, but the vendor evidence in this page's sources is claims without published accuracy. If you do buy detection, ask the detection software evaluation questions first.
One structural option is an AI interview platform that runs autonomous interviews, scores each answer with written reasoning and keeps a timestamped transcript. That puts the evidence in front of a human reviewer; it does not detect AI assistance.
What is the legal and fairness frame?
Article 5(1)(f) of the EU AI Act prohibits AI systems that infer a person's emotions in the workplace, except for medical or safety reasons, from 2 Feb 2025, so AI that infers a candidate's emotions from facial expressions is widely read as prohibited in EU hiring, though the cited text does not mention recruitment or lie detection (AI Act, Art. 5).
The Illinois AI Video Interview Act and BIPA also bear on video interviews; this page does not cover them. NYC Local Law 144 obligations may apply where monitoring influences decisions.
Accommodations matter in practice. Gaze, latency and pacing can differ for disabled and neurodivergent candidates, and each of those is also something a vendor might count as a sign of AI help. Any flag should go to a human review, and candidates need a way to ask for an adjustment.
Copy-ready candidate AI-use statements
Each statement is plain and non-accusatory, covers application, take-home and live stages, names the consequence in neutral terms, and gives candidates a route to ask for an accommodation. Adapt them with counsel.
Prohibit live use
You are welcome to use AI tools to prepare and to draft your application. For take-home exercises and live interviews, please do not use AI tools or other assistance unless we say so in the invitation. If we have a concern during the process, we will talk with you before any decision. If you need an adjustment, tell us in advance and we will arrange it.
Allow and disclose
You may use AI tools in the application, the take-home exercise and the live interview. Please tell us at the start which tools you are using and how. We will ask you to explain your answers in your own words. If you need an adjustment or a different format, tell us in advance and we will arrange it.
Structure the interview around AI
Preparation with AI is welcome, including for your application and any take-home exercise. In the live interview, we will ask you to explain and extend your answers, and some questions will be new to you. Our aim is to see your own reasoning. If a format does not work for you, tell us before the interview and we will find another.
FAQ
Is using ChatGPT in a job interview allowed?
It depends on the employer's policy. Anthropic, for example, says live interviews are "all you–no AI assistance unless we indicate otherwise" (Anthropic). If the invitation does not say, ask the recruiter.
Do recruiters know when you use ChatGPT?
Usually not in a live conversation. In interviewing.io's test, none of 32 analysed interviewers suspected ChatGPT help. The sample is small and the test is undated, so treat it as a direction rather than a rate.
How do you test a candidate who uses AI?
Use unscripted follow-ups on the candidate's own answers, custom problems and judgment questions. Custom questions cut the cheater pass rate from 73% to 25% in the interviewing.io test.
Should we let candidates use AI in interviews?
There are three defensible options: prohibit live use, allow with disclosure, or structure the interview around AI. Decide by what you are measuring. The decision matrix above sets out the trade-offs.
What should an AI-use policy say?
State what is allowed at each stage (application, take-home and live interview), what candidates should disclose, how to request an accommodation, and what happens if there is a concern. Use neutral wording for the consequence.
Can AI-text detectors catch interview answers?
Not reliably. One source says such tools "don't work any more," and research on essays found that detectors can misclassify non-native English writing. They are not a control.
Is preparing with AI before an interview cheating?
No. Preparation is not live assistance. Where policies differ, it is on take-home work and live interviews, not on preparation.
Does a candidate using AI count as interview fraud?
It is distinct from impersonation. Whether AI help is a violation depends on the policy the candidate was given, so the policy question comes first.
Next step
A ZeaHire evidence record contains the transcript, per-question scores and proctoring counters for a human reviewer. See what a ZeaHire evidence record contains.
Comments
No comments yet — start the conversation.