TL;DR: Interview cheating detection software checks presence and environment (faces, phones, gaze), answer-content patterns (response timing, coached phrasing, tool fingerprints) and, in some products, identity (ID and selfie match). No vendor publishes false-positive rates, so a flag is a reason to review, not proof. Choose by what the tool shows a human reviewer for each flag, what it does automatically, and what it stores.
Updated 28 October 2026. Facts last verified 15 Oct 2026.
Key takeaways
- Gartner's survey (n=3,000 candidates) found 6% admitted to interview fraud (Gartner, 31 Jul 2025). Greenhouse's November 2025 report (n=4,136) found 91% of U.S. recruiters surveyed had spotted candidate deception (Greenhouse, 19 Nov 2025).
- Fabric, a vendor, states in its own published guidance that no vendor publishes false-positive rates. "Flagged" is not "confirmed."
- Gaze, speech-pattern and AI-text flags can mis-flag candidates, including non-native-English writers. The study of AI-text detectors (Liang et al., 2023) tested writing samples, not interviews.
- Monitoring candidates triggers legal duties. The EU AI Act bans emotion-inference AI in the workplace from 2 Feb 2025, biometric data processed to identify a person is special-category data under GDPR Article 9, and NYC Local Law 144 may apply when a flag feeds a score. Take legal advice on each point.
- Redesigning the interview often does more than detection, which is the conclusion of a vendor source (Fabric). TheHireHub, another vendor-affiliated source, adds that publishing AI-use rules helps (TheHireHub).
Employers evaluating detection software find product pages that list what a tool looks at and say little about how often it is wrong. The vendors' own guidance concedes the gap: Fabric states that no vendor publishes false-positive rates. This page sets out what the tools check, how reliable each signal is, and how to choose between products.
This page is the buyer's guide to the tooling layer. The record of what has happened, and the layers of defence around the interview, are covered in interview fraud. Neighbouring topics get one paragraph each, then a link.
How we compiled this page. We checked each source on 15 Oct 2026 unless a date is given. Criteria: a primary source (publisher, regulator, court or agency) where one exists; a vendor's own page, labelled as the vendor's, where it does not. Terms: Published means the information appears on a page we read. Vendor-claimed means a vendor states it and no independent test was found. Not published means we found no statement. We list no product as inaccurate.
What is interview cheating detection software?
Interview cheating detection software is a set of tools that flag signals suggesting a candidate received unauthorised help or is not the person they claim to be. The signals fall into three groups: presence and environment (faces, phones, gaze), answer content (timing, coached phrasing, tool fingerprints) and identity (ID document and selfie match, liveness).
This page is for employers evaluating tools. It does not describe evasion. It uses "flagged for review" for a candidate and avoids "cheater."
How common is the problem it addresses?
The prevalence evidence comes from surveys with different samples and different definitions, so the figures should not be added together.
| Source | Date | Sample | Finding | Type |
|---|---|---|---|---|
| Gartner, via HR Dive (8 Aug 2025) | 31 Jul 2025 | 3,000 candidates | 6% admitted interview fraud. Gartner predicts 1 in 4 candidate profiles worldwide will be fake by 2028 (a prediction). | Analyst survey, self-reported |
| Greenhouse AI in Hiring Report | 19 Nov 2025 | 4,136 respondents, including 1,236 recruiters and hiring managers | 91% of U.S. recruiters surveyed (U.S. n=665 recruiters and hiring managers) had spotted candidate deception. Among hiring managers, 65% caught deceptive AI use (32% AI-generated scripts, 22% hidden prompt injections in resumes, 18% deepfakes). 74% were more concerned than a year earlier. 46% of U.S. job seekers said trust in hiring had fallen. | Vendor survey |
Impersonation is also a security issue. The FBI's Internet Crime Complaint Center (28 Jun 2022) reported complaints of deepfakes and stolen personal information used in applications for remote IT and database roles, and listed audio cues, such as coughing or sneezing that does not match what is visible, as tell-tale signs. On 30 Jun 2025 the US Department of Justice announced coordinated actions against North Korean remote IT worker schemes, describing 100+ US companies affected and searches of laptop farms.
What do these tools actually look at?
The tools check three groups of signals.
Presence and environment. Multiple people in frame, a missing face, eye and head tracking, screen overlap and phone detection. Alex's Cheat Detection add-on is a documented example. We fetched its documentation on 15 Oct 2026: the add-on lists multiple-people analysis, eye and head tracking, voice analysis for external voices or voice switching, and transcript analysis. It returns a three-level result (Unlikely, Possible or Likely) with per-analysis scores, and routes flagged interviews to the team "without making definitive claims."
Answer content. Coached-answer patterns, response timing, AI-tool fingerprints and voice authentication. Fabric ranks coached-answer patterns and tool fingerprints above proctoring and voice as more reliable. That is a vendor ranking, and we report it as one. For what the sellers of real-time assistants claim about themselves, see what real-time interview assistants claim.
Identity. Identity checks are the third group, such as ID document and selfie match, liveness checks, and continuity of face or voice across stages. We cite no product for them here, so we make no claim about how any tool performs them.
Documented examples fall into two groups. We name only products for which we hold a primary source and a date.
| Group | Documented example | Source and date |
|---|---|---|
| Autonomous AI interviewers with integrity add-ons | Alex; Fabric | Alex documentation (15 Oct 2026); Fabric (28 Aug 2026, modified 10 Oct 2026) |
| Academic work | Cheating Detection Pipeline for Online Interviews and Exams (arXiv, 2021) | We have not extracted a summary of this paper. |
How reliable are the signals? (the false-positive thesis)
No vendor publishes false-positive rates. Fabric states this in its own guidance and adds that a published rate "becomes a target." We write "not published" for every tool and do not describe any product as inaccurate.
The matrix below sets out what each signal catches, why it can mislead, where regulation applies and what corroboration to require. It is our assessment built from the sources in this section. It does not contain measured error rates, because none has been published.
| Signal | What it catches | Typical false-positive causes | Jurisdictional risk | Corroboration to require |
|---|---|---|---|---|
| Multiple faces | Another person in frame | Shared rooms, household members, reflections | Low | Review frame and timestamp; ask the candidate |
| Gaze and head tracking | Looking away from the screen; possible reading | Neurodivergence, disability, camera position, thinking pauses | EU: must not infer emotion (AI Act Art. 5(1)(f)); biometric rules if face data is stored | Review the video; compare with answer content |
| Phone or object detection | A phone or object in frame | Permitted use of a phone for the interview link or sign-in | Low | Review frame; ask the candidate |
| Voice switch | A second voice or a change of voice | Shared rooms, background audio, audio compression, accent | Voice matching is biometric data; GDPR Art. 9 may apply, depending on the tool | Listen to the segment; compare with other rounds |
| Response latency | Uniform or delayed answer timing | Connection lag, thinking time, language processing time, disability | NYC Local Law 144 may apply if it feeds a score | Never act on timing alone |
| Coached-answer patterns | Phrasing matching scripted or generated answers | Well-prepared candidates, structured speakers, rehearsed phrasing | NYC Local Law 144 may apply if it feeds a score | Ask a follow-up on the candidate's own answer |
| AI-text detector | Text likely produced by a language model | Non-native English writing, as tested on essays by Liang et al. | NYC Local Law 144 may apply if it feeds a score | Independent evidence; never the detector alone |
| ID plus selfie | Whether the face on an ID matches the live person | Lighting, camera quality, appearance change, worn documents | Biometric data: GDPR Art. 9 and Illinois BIPA may apply, depending on the tool | Manual review; an alternative route when a match fails |
| Liveness | Whether a live person is present | Low light, low-end cameras, assistive devices | Biometric rules if face data is stored | Manual review; an alternative verification route |
| Device or session | Unusual device or session changes | Shared computers, network changes, corporate proxies | Data-protection rules on device identifiers | Cross-check with the candidate's explanation |
Detection of AI-written text is a known weak area. Liang et al., "GPT detectors are biased against non-native English writers" (arXiv 2304.02819, submitted 6 Apr 2023, final version 10 Jul 2023), found that detectors consistently misclassified non-native English writing as AI-generated. The study tested writing samples, not interview video.
Gaze and speech-pattern flags can also mis-flag candidates whose natural behaviour differs from the norm. FloCareer's limitations section (15 Apr 2026) lists false positives from natural behaviour and privacy limits. Both TheHireHub and Alex's documentation say a human should review each flag. The red flags that need a human check give the corroboration steps to require before any adverse action.
Detection or design?
Fabric and TheHireHub both argue that redesigning the interview outperforms detection alone: judgment questions, dynamic follow-ups and an identity check at the first live stage. TheHireHub also lists published AI-use rules. Both are vendor-affiliated, so we use their argument as a structure and not as evidence. Fabric's guidance carries the heading "Detection Is the Second-Best Answer."
The candidate AI-use policy question, with the full matrix of options, is covered on its own page and is not repeated here.
TheHireHub's five-step sequence (TheHireHub, 18 Aug 2026):
- Publish AI-use rules.
- Verify identity once, at the first live stage.
- Redesign questions toward judgment over recall.
- Route flags to named human review.
- Require corroborating evidence before rejecting.
Legal guardrails for monitoring candidates
This section is a summary of sources, not legal advice.
EU AI Act. Article 5(1)(f) bans AI systems that infer emotions in the workplace (with medical and safety exceptions) from 2 Feb 2025 (EU AI Act, Article 5). Detection features that infer stress or emotion are a red flag for EU use; see the EU AI Act workplace emotion-inference ban for the recruitment rules. Under Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026, the high-risk obligations for Annex III systems, which include employment, apply from 2 December 2027.
GDPR. Biometric data processed to uniquely identify a person is special-category data under GDPR Article 9. Matching faces or voices can fall within it. What applies depends on what each tool does, and we have not checked this for any product. The UK Information Commissioner's Office reported on 6 Nov 2024 that its audit of AI recruitment tools found over-collection of personal data and indefinite retention.
United States. Scoring or ranking candidates with an automated tool can trigger the NYC bias-audit rules: a bias audit, and notice at least 10 business days before use, enforced from 5 Jul 2023 (NYC Department of Consumer and Worker Protection). Whether an integrity flag that feeds a score is itself an output of an automated employment decision tool is a question for counsel. Whether a given tool's face processing falls under Illinois BIPA is also a question for counsel; check the statute and case law.
Ten questions to ask any detection vendor
Each question should be answerable yes, no or not published.
- Does the recruiter see the timestamp, the signal and the underlying audio or transcript for every flag?
- Is a flag ever an automatic rejection or score deduction?
- Is there an appeal path for the candidate?
- What does the vendor state about false positives and demographic testing?
- What identity data is stored, for how long, and where?
- Does the tool infer emotion?
- Which meeting platforms does it support?
- What is the candidate disclosure text?
- Are detections logged for audit?
- What does the tool do when a candidate requests an accommodation?
On candidate experience, intrusive methods such as full room scans or extensive biometric tracking raise privacy issues and can hurt the candidate experience (FloCareer). Greenhouse found that 46% of U.S. job seekers said trust in hiring had fallen over the past year, so a proportionate, disclosed approach matters (Greenhouse, 19 Nov 2025).
Where integrity features sit in AI-interview roundups
For how integrity features compare across AI-interview products, see best AI interview software.
When comparing platforms, weigh integrity features as one criterion among the others: what is shown to the reviewer for each flag, what happens automatically, what is stored and for how long, and whether the disclosure to candidates is clear. Integrity features differ in kind between autonomous interviews, where the tool runs the session, and live-meeting add-ons, which observe a call the tool does not control; see AI interview tools for the category.
FAQ
How does AI detect cheating in interviews?
AI tools check three groups of signals: presence and environment, answer content and, in some products, identity. These are vendor-reported, and no vendor publishes accuracy figures.
Can AI interviewers detect cheating?
AI interviewers can log signals for a human to review, but none publishes false-positive rates. A logged signal is a lead, not a finding.
How accurate is cheating detection?
Accuracy is not published by vendors. Fabric states this itself in its guidance, and we found no vendor that contradicts it.
Does detection software create false positives?
Yes, false positives are the central risk. AI-text detectors have a documented weakness with non-native English writing (Liang et al., tested on essays), and gaze and speech flags can mis-flag some candidates.
Is it legal to monitor candidates during interviews?
It depends on the jurisdiction. The EU AI Act bans emotion inference in the workplace, biometric matching can trigger GDPR Article 9 duties, and NYC Local Law 144 may apply when a flag feeds a score. Take advice from counsel before deploying any tool.
What do you do when a candidate is flagged?
Send the flag to a named human reviewer and require corroborating evidence. Add a verification step, and do not use rejection language until that step is complete.
Does a "flag" mean a candidate cheated?
No. A flag means the candidate is flagged for review, and nothing more.
Should we buy detection or redesign the interview?
Both are options, and the vendors cited here say redesign comes first. Fabric and TheHireHub both argue that interview design, and per TheHireHub published AI-use rules, outperform detection alone.
Comments
No comments yet.
Comments are closed on this post.