TL;DR: In AI-coached interview cheating, sellers of real-time AI interview assistants advertise that screen sharing cannot see them. That is the sellers' own claim, not an independent finding. Fabric states that no vendor publishes false-positive rates for detection, so a flag is a reason to review, not proof. The defences with evidence are interview design and a published AI-use policy, not screen-share rules.
Updated 28 October 2026; tool claims re-checked 15 Oct 2026.
Key takeaways
- Cluely's homepage describes itself as "Invisible to screen share" and Final Round AI's site says its windows are "Hidden from shared screens." Both are marketing claims, fetched 15 Oct 2026 (Cluely, Final Round AI), and independently unverified.
- In interviewing.io's test (37 interviews, 32 analysed), no interviewer suspected AI help (interviewing.io).
- In that test, custom questions cut the pass rate of participants using ChatGPT from 73% (verbatim questions) to 25%. Anthropic Engineering reports redesigning its own take-home repeatedly as models improved (Anthropic Engineering, 21 Jan 2026).
- Vendors say behavioural signals (response timing, reading gaze, cadence) can flag use, but Fabric's own page states that no vendor publishes false-positive rates. A flag is not proof.
- Anthropic's candidate guidance shows a simple policy: no AI assistance in live interviews unless the employer says otherwise.
Detecting AI-assisted answers is a live concern for hiring teams. In Checkr's survey of 3,000 US managers, published 16 Sep 2025, 59% suspected AI-based misrepresentation by candidates and only 19% were extremely confident of catching fraud. Sellers of real-time assistants advertise that screen sharing cannot see them, while vendors of detection say behavioural signals can flag the use anyway.
This page separates what the tools claim from what has been tested, then sets out the defences that have evidence behind them. It covers one form of the wider picture of interview fraud, which holds the sourced record of cases and the layers of defence.
What is AI-coached interview cheating?
AI-coached interview cheating is the use of a real-time assistant to generate or shape a candidate's answers during an interview, without the interviewer's knowledge. Seller pages below describe two channels: a screen overlay that displays suggested answers, and a human helper who sends notes as the interview proceeds. In our reading, two more are common: audio capture that passes the conversation to a model, and a second device running a chatbot. We found no seller page for those two, so treat them as our description rather than a documented product feature.
It is not the same as impersonation, where someone else answers in the candidate's place. It is also not the same as allowed preparation, such as practising answers beforehand or using notes where the employer permits them. Whether a particular use of an assistant counts as cheating depends on the employer's stated policy, which the defences section returns to.
What do the tools claim?
Everything in this section comes from the sellers' own pages, fetched 15 Oct 2026. These are marketing claims. They have not been independently verified, and this page does not recommend any product.
- Cluely. The homepage calls the product "Undetectable AI for Meetings" and lists "Invisible to screen share" and "300ms Response time."
- Final Round AI. The site describes a Stealth Mode with windows "Hidden from shared screens" and "Absent from recordings" and states the product is "100% invisible on Zoom, Google Meet, Teams, HackerRank, CoderPad."
- LockedIn AI. Its FAQ describes "LockedIn Duo," where a helper joins a separate private session and sends notes in real time, and says the arrangement is "invisible to the interviewer" because it sits outside the interview call. This is a human-in-the-loop variant: the coaching comes from a person, not only a model.
One dated fact about Cluely's origin is on the record. TechCrunch reported on 5 Mar 2026 that the company was founded after its CEO described his suspension from Columbia University for building interview-cheating tools.
Interview Coder and Parakeet AI also appear in search results for these queries. Their capabilities have not been verified, so this page does not describe them. Sellers and reviewers also describe the underlying mechanisms of overlays and audio capture; those accounts come from third-party review blogs, are unverified, and are left out here.
Does screen sharing stop AI cheating?
No rule about screen sharing settles the question. The sellers' own headline claim is that their windows are hidden from shared screens, and no independent test confirming or refuting that was found. Interview structure and a published policy are the more reliable response.
Prevention advice often starts with the share itself. One recruiter guide, published by Alex on 8 Oct 2025, lists full-screen sharing among its prevention steps. That step addresses what a shared screen displays. It does not address the sellers' claim that the assistant never appears on the shared screen, nor a second device or a remote helper, which sit outside the share entirely.
Claim, surface and counter (the table)
The table below sets each seller's claim beside the channel it uses, what a screen-share rule does about it, which signals vendors say could apply, which design counter is available, and how strong the evidence is.
| Tool or setup | What the seller claims (fetch date) | Channel | What screen-share rules do about it | Signals that could apply (vendor-asserted) | Design counter | Evidence grade |
|---|---|---|---|---|---|---|
| Cluely | "Undetectable AI for Meetings"; "Invisible to screen share"; "300ms Response time" (cluely.com, 15 Oct 2026) | Screen overlay and audio capture, as sellers and reviewers describe | Nothing, if the claim holds: the seller says the window is invisible to a screen share | Response timing, reading gaze | Custom problems; unscripted follow-up on the candidate's own answer | Vendor claim; no independent test found |
| Final Round AI | "100% invisible on Zoom, Google Meet, Teams, HackerRank, CoderPad" (finalroundai.com, 15 Oct 2026) | Screen overlay ("Stealth Mode" windows) | Nothing, if the claim holds | Response timing, reading gaze | Custom problems; short live-reasoning segment | Vendor claim; no independent test found |
| LockedIn Duo (remote helper) | Helper joins a private session and sends notes in real time; "invisible to the interviewer" (lockedinai.com, 15 Oct 2026) | Human helper | Nothing: the helper is outside the interview call, per the seller's description | Response timing, speech cadence | Unscripted follow-up on the candidate's own answer; in-person round where proportionate | Vendor claim; no independent test found |
| ChatGPT on a second device | None (general-purpose chatbot, not a seller claim) | Second device | Nothing: a second device is not part of the shared screen | Response timing, reading gaze | Custom questions (73% to 25% pass rate in interviewing.io's test) | Primary, one small test of the counter (undated) |
Evidence grade key: Primary is an independent or first-party record. Vendor is a seller's or vendor's own statement. None means nothing was found. Design counters are editorial except the custom-question result, which comes from interviewing.io's test. No counter has been tested against these specific products.
How big is it, and how good is the evidence?
The evidence base is thin, and it comes mostly from vendors and small tests.
- Fabric (vendor, own platform). Fabric reports 19,368 AI-led interviews between Jul 2025 and Jan 2026, with 38.5% flagged, a flagged rate of about 48% in engineering roles (12% in sales), and incidents that roughly tripled in late 2025 (Fabric report). A flag is not confirmed cheating. Fabric's separate detection page, modified 10 Oct 2026, states that no vendor publishes false-positive rates.
- Gartner (independent survey). A 2Q25 Gartner survey of 3,000 job candidates found 6% admitted to participating in interview fraud, and Gartner predicts that by 2028 one in four candidate profiles worldwide will be fake (Gartner, 31 Jul 2025). Gartner's survey measures admitted fraud in general, not coached cheating specifically.
- Checkr (survey of 3,000 US managers, 16 Sep 2025). 59% suspected AI-based misrepresentation by candidates and 19% were extremely confident of catching fraud (Checkr). Checkr is a vendor and this is a vendor survey.
- interviewing.io (undated, small). Of 37 runs, 32 were analysed, and the interviewer suspected ChatGPT help in none of the 32 (interviewing.io).
Claims about prevalence should be held loosely: the figures above are a vendor's own flag rate, a vendor's survey, a self-reported survey and a small test.
Which signals do vendors say they use, and what are the limits?
Vendors of detection say several behavioural signals can indicate assistant use: flat response timing, reading gaze and answers that do not hold up under unscripted follow-ups, according to Fabric. Fabric's guidance also describes fingerprint detection based on browser tab state, clipboard behaviour, injected overlay elements and audio side-channels. These are vendor assertions. No independent audit of any of them was found, and no vendor publishes false-positive rates.
A buyer comparing products against criteria can use how to evaluate detection software, which sets out what to ask each vendor.
One legal limit applies to how these signals can be used. The EU AI Act limits on emotion inference rest on Article 5(1)(f), which bars emotion inference in the workplace from 2 Feb 2025 (text of Article 5). Eye tracking should therefore be limited to presence or attention logging, not used to infer deception, and reviewed by counsel before use.
Language matters as well. A signal leaves a candidate "flagged for review", never labelled a "cheater". A flag goes to a human reviewer, and corroborating evidence is required before any adverse action.
Defences that have evidence: design and policy
Two sources give measured or documented results on design.
- Custom questions. In interviewing.io's test, replacing verbatim questions with custom ones cut the pass rate of participants using ChatGPT from 73% to 25% (interviewing.io). The control pass rate was 53%. The test was small and is undated.
- Out-of-distribution tasks. Anthropic Engineering (Tristan Hume, 21 Jan 2026) describes redesigning its take-home as models improved: tasks that sit outside what a model has seen, and candidates who must build their own tooling (Anthropic Engineering).
Three further practices are design recommendations rather than tested results: unscripted follow-ups on the candidate's own answer (the structured interview question bank gives examples), a short live-reasoning segment, and an in-person round where it is proportionate to the role.
Interview formats that score each answer with written reasoning and keep a timestamped transcript give a reviewer something to check. That is a property of evidence-based AI interviews, not a detection claim.
Policy is the second defence. Anthropic's candidate guidance, updated 10 Jul 2025, tells candidates that live interviews are "all you–no AI assistance unless we indicate otherwise."
Choosing between an allow, prohibit or redesign policy is set out in full on its own page; this page does not repeat it.
What to do if a candidate is flagged
- Pause. Do not act on the flag alone.
- Verify through a second channel. Confirm the point in question by a separate conversation or task.
- Document. Record what was flagged, when, and what was checked.
- Involve HR before any decision is made.
Do not accuse the candidate in the interview. Rule out innocent explanations first: nerves, notes the employer allows, a disability, or a poor connection.
The verification steps are in spotting fake candidates in video interviews, and are not repeated here.
FAQ
Is Cluely detectable?
The seller claims it is invisible to screen share, and no independent test of that claim was found. ZeaHire makes no detection claim about Cluely or any other tool.
Does screen sharing stop AI cheating?
No rule about screen sharing settles the question. The sellers' own headline claim is that their windows are hidden from shared screens, and no independent test confirming or refuting that was found. Interview structure and a published policy are the more reliable response.
Can Zoom detect Cluely?
No primary source was found that answers this, so this page does not assert either way. The claim-versus-evidence table above shows what the seller says and how strong the evidence is.
How do interviewers catch AI answers?
In interviewing.io's small test, no interviewer suspected AI help in the 32 interviews analysed. Interview design, unscripted follow-ups and a published policy help more than hoping to notice.
Is using an AI interview copilot cheating?
It depends on the employer's stated policy. Anthropic's rule is one example: live interviews are "all you–no AI assistance unless we indicate otherwise."
Do AI interview cheating tools work?
No independent evidence either way was found. The sellers' claims are marketing statements, and the only test results cited here concern interview design.
How do you make interviews cheat-resistant?
Use custom problems, unscripted follow-ups on the candidate's own answer, a short live-reasoning segment and, where proportionate, an in-person round. Publish the AI-use policy before the interview.
Does ZeaHire detect real-time AI help?
In autonomous AI interviews, ZeaHire's proctoring report counts events such as look-aways and suspected reading. It does not claim to detect named tools or overlays.
Next step
See what a ZeaHire evidence record contains: transcript, per-question scores and proctoring counters, for a human to review.
Comments
No comments yet — start the conversation.