ARTICLE

Cluely Detection: What Employers Can Really Do About AI Copilots

Cluely detection is not possible from a browser. Learn what employers can actually do about AI interview copilots: test design, timing signals and fair review.

Cluely Detection: What Employers Can Really Do About AI Copilots

TL;DR

Invisible AI copilots like Cluely cannot be seen by a browser, so detection alone will never settle the question. This guide shows small employers how test design, timing signals and fair human review can reduce the problem without overclaiming.

Cluely detection is not something any website or browser-based test can do reliably. Invisible AI interview copilots run as separate desktop apps, outside the browser's view. Employers get further with test design that leaves no time to read an answer back, plus behavioral signals such as timing anomalies and focus loss, reviewed by a person.

That answer disappoints some hiring managers, but it is the honest one. Tools such as Cluely and Interview Coder are marketed as invisible during screen sharing, and many small businesses now wonder whether their online tests and video interviews still mean anything. They can, if you change what you rely on.

This guide explains how an AI copilot works, what a browser can and cannot observe, how to design assessments that make real-time AI help hard to use, which signals deserve a closer look, and how to respond fairly without accusing anyone.

What are AI interview copilots like Cluely and Interview Coder?

An AI interview copilot is software that listens to or reads an interview or test in real time, sends the content to a large language model, and displays a suggested answer to the candidate. Cluely and Interview Coder are two widely discussed examples. Interview Coder was promoted for technical coding interviews, while Cluely has been pitched more broadly for meetings, calls and interviews.

How an overlay copilot works

Most of these tools follow the same basic loop. The app captures what is on the screen or what the interviewer says, often through a screenshot or audio transcription. It sends that content to an AI model, receives a response within a few seconds, and shows the response in a transparent window layered over other applications.

The key design goal is that the overlay window does not appear when the candidate shares their screen and does not take keyboard focus away from the interview or test. From the browser's point of view, the candidate never left the page.

Why AI copilots matter for hiring tests, not just live interviews

Most coverage of AI interview cheating focuses on live video calls. Small businesses face the same issue on take-home and online pre-employment tests, because the same overlay can sit on top of a test tab as easily as a video call. Any assessment that a candidate completes unsupervised on their own computer is exposed to an AI copilot in some form.

That does not make online testing pointless. It means the test has to be built so that a copilot provides less advantage than it promises.

Can employers detect Cluely or Interview Coder?

No browser-based test or video platform can reliably detect an invisible overlay copilot running as a separate desktop app. Browsers are sandboxed for security, so a web page cannot list the other programs on a computer or see windows drawn above it. A copilot on a second device, such as a phone, is invisible to the computer entirely.

Any vendor that promises Cluely detection or interview coder detection from inside a browser tab is overclaiming. What a browser can do is observe how a candidate interacts with the test itself, and those observations are useful evidence when read carefully.

What a browser can see

A browser-based assessment can record events that happen inside its own page. Common examples include:

  • The tab or window losing focus, such as a switch to another tab or app
  • Paste events and blocked copy attempts
  • Time spent on each question, measured by the server
  • Typing rhythm, such as long silences followed by bursts of text
  • Exiting full-screen mode, where the test requests it

What a browser cannot see

A browser cannot see other installed software, transparent overlays, audio captured by another app, a phone propped beside the keyboard, or a person in the room. Well-built overlay copilots are designed specifically so they do not trigger focus loss. So the absence of integrity signals proves nothing about whether help was used, and their presence proves nothing about cheating either.

BehaviorVisible to a browser test?What it can reasonably suggest
Switching tabs or appsYes, as focus lossThe candidate left the test; reason unknown
Pasting text into an answerYesText came from somewhere else
Unusual timing per questionYes, if timed on the serverA pattern worth reviewing, not a conclusion
Invisible overlay copilotNoNothing directly
AI app on a phoneNoNothing directly
Someone helping off cameraNo, unless a camera check is used with consentCamera images may show it, for a person to judge

Why test design beats detection

Since detection has hard limits, the stronger strategy is to make an AI copilot slow, awkward or unhelpful to use. You cannot block the tool, but you can shrink the window in which it is useful.

The read-back window

Every copilot answer passes through four steps: capture the question, wait for the model, read the suggestion, then enter or say it. Call the total time the read-back window. On an untimed multiple-choice quiz, that window is irrelevant, because the candidate has unlimited time. On a well-designed timed task, the window eats into the time the candidate needs to actually complete the work.

The practical insight is this: copilots are strongest on questions where the answer is short and the task ends when the answer is known. They are weakest on tasks where knowing the answer is only the start, and the candidate still has to do several steps of interaction to show it.

Design choices that shrink the read-back window

These choices make AI assistance harder to use without making the test unfair to honest candidates:

  1. Time each question separately. A per-question timer set for a capable person, not a generous overall timer, leaves little spare time for capture, waiting and reading.
  2. Serve one question at a time from the server. Candidates cannot screenshot the whole test at the start and process it in advance.
  3. Give each candidate a different version. Randomized items and shuffled values stop answers from being shared or memorized.
  4. Favor doing over knowing. Spreadsheet tasks, typing tests, data entry and drag-and-drop sequencing require hands-on work that a suggested answer does not complete for the candidate.
  5. Use multi-step problems with intermediate inputs. Computational thinking and problem-solving tasks that ask for each step, not just the final number, make a single pasted answer less useful.
  6. Ask for judgment tied to your business. Short written answers about a realistic scenario from your own workplace are harder to answer well with generic AI text, especially when scored against your own rubric.

For more tactics that apply to every kind of AI assistance, not only overlays, see this guide on how to reduce AI assistance in online assessments.

Example: a small accounting firm hiring a bookkeeper

Imagine a five-person accounting firm hiring a remote bookkeeper. Its first test was 30 untimed multiple-choice questions about accounting terms. Every candidate scored above 90 percent, which told the owner almost nothing.

The redesigned assessment uses a timed spreadsheet task to reconcile a short ledger, a numerical reasoning section with per-question timers, and one written answer about how the candidate would handle a client who sent incomplete receipts. A copilot can still suggest definitions, but it cannot fix the ledger cells for the candidate inside the time limit. Scores now spread out, and the owner has something meaningful to discuss in the interview.

Signals worth reviewing: timing anomalies and focus loss

Behavioral signals cannot see a copilot, but they can show patterns that are worth a conversation. Treat every signal as a reason to look closer, never as a finding on its own.

Timing anomalies

Timing anomalies are response patterns that do not match how people usually work through a test. Examples include hard questions answered as fast as easy ones, nearly identical times across every question, or a long pause on each item followed by a quick, confident answer. A long pause then a fast answer can match the read-back window of a copilot. It can also match a careful person who thinks before clicking.

Focus loss and pastes

Focus loss means the test window stopped being the active window. Repeated focus loss just before strong answers is a pattern to note. A single focus loss usually means a notification, a dropped connection or a child asking for a snack. Pastes into a written answer are clearer evidence that text came from elsewhere, but not who wrote it.

What integrity signals cannot tell you

Signals never tell you intent. Candidates with disabilities may use screen readers, magnifiers or extra time, which change timing and focus behavior. Candidates on older laptops or unstable connections produce noisy data. If you grant accommodations, read signals in light of them. The guide to pre-employment test accommodations covers how to offer them consistently.

For a deeper look at reading these patterns across many kinds of tests, see how to detect AI cheating in online hiring tests.

Live interviews: what to do when an AI copilot might be on screen

In a live interview, the best defense against an AI interview copilot is conversation that adapts in real time. Scripted questions are easy to feed to a model; follow-up questions about the candidate's own previous answer are much harder.

Useful habits for interviewers include:

  1. Ask why, then ask what if. After an answer, ask the candidate to explain their reasoning, then change one condition and ask how the answer changes.
  2. Build on their test results. Pick a question the candidate answered well and ask them to walk through it. Someone who did the work can usually reconstruct their thinking.
  3. Ask for specifics from experience. Concrete details about a past project, the tools used and what went wrong are harder to generate convincingly on the spot.
  4. Notice rhythm, not just content. Very polished answers paired with an unchanging delay before every response can be worth a gentle follow-up, though nerves and second-language speakers can show similar patterns.
  5. State the rules at the start. Tell candidates plainly whether outside tools are allowed. Clear expectations are fairer than silent suspicion.

Never confront a candidate with an accusation based on a hunch. If something feels off, gather more evidence through a structured follow-up task.

A practical response plan for small businesses

Without a recruiting team, you need a simple, repeatable process. This five-step plan keeps decisions fair and consistent:

  1. Publish your rules. Tell every candidate in writing what help is and is not allowed during tests and interviews.
  2. Design for the read-back window. Use timed, hands-on and multi-step tasks for the skills that matter most in the role.
  3. Collect signals, with consent where cameras are involved. Explain any monitoring up front. Some states regulate biometric data, such as Illinois under its Biometric Information Privacy Act, so check with a qualified advisor before adding face or identity checks.
  4. Review, do not reject automatically. A person should look at scores and signals together, alongside the resume and interview, before any decision.
  5. Verify with a second step. When results and signals do not fit together, offer a short live follow-up or a supervised task to every candidate in the same situation.

Fair process matters legally as well as ethically. The U.S. Equal Employment Opportunity Commission (EEOC) enforces federal laws such as Title VII of the Civil Rights Act and the Americans with Disabilities Act, which apply to employment tests and selection procedures. The EEOC's guidance on employment tests and selection procedures explains the general principles. For your specific situation, talk to an employment attorney or HR advisor.

Limitations and trade-offs

No approach removes AI interview cheating entirely, and it helps to be clear about the costs of each choice.

  • Tight timers reduce copilot advantage but can disadvantage slower readers and people who need accommodations, so offer extra time when requested.
  • Camera checks add evidence but feel intrusive to some candidates and raise privacy obligations.
  • Hands-on tasks take longer to build and complete, which can lower completion rates for entry-level roles.
  • Signals produce false alarms, so they cost reviewer time and require judgment.
  • Live follow-ups are the most convincing check but do not scale to dozens of applicants.

A sensible balance for most small businesses is a short, timed, hands-on online assessment for everyone, followed by a structured live conversation for the shortlist.

Frequently asked questions

Is using an AI copilot in a job interview illegal?

Using an AI copilot is generally not a crime, but it may break the rules an employer sets for the hiring process, and misrepresenting your abilities can lead to a withdrawn offer. Employers should state their rules clearly in writing before testing or interviewing. Laws vary by state and situation, so consult a qualified advisor for anything specific.

Can a webcam detect Cluely?

A webcam cannot see software running on a screen, so it cannot detect Cluely directly. Camera images may show a candidate repeatedly reading something off to the side or a second person in the room, which a reviewer can consider. Any camera check should be disclosed to candidates in advance and used only with their consent.

Does full-screen mode stop overlay copilots?

Full-screen mode does not stop overlay copilots, because overlays are drawn above other windows by the operating system. Full-screen mode does make tab switching more visible, since leaving it can be recorded. It is a useful signal for ordinary tab switching, but it is not a defense against tools designed to stay invisible.

Should I allow AI tools on some pre-employment tests?

Allowing AI tools can make sense when the job itself involves using them, such as drafting marketing copy with AI assistance. In that case, test how well candidates direct, check and edit AI output. Be explicit about which sections permit AI and which do not, so candidates are judged against the same rules.

Do proctoring tools prove that a candidate cheated?

Proctoring tools and integrity signals do not prove cheating. They record observations, such as focus loss, pastes or unusual timing, that a person must interpret in context. Technical glitches, disabilities, accommodations and home distractions can all produce similar patterns. Use signals to decide where to look closer, then verify with a follow-up step offered fairly.

What types of tests are hardest for AI copilots to help with?

Timed, hands-on tasks are hardest for AI copilots, including spreadsheet exercises, typing and data entry, multi-step problem solving with intermediate answers, and written responses about scenarios specific to your business. These tasks require the candidate to act, not just know an answer, so a suggestion on screen provides much less advantage.

How often should I update my hiring tests?

Review your hiring tests at least once a year, and sooner if scores suddenly cluster at the top or no longer predict on-the-job performance. Rotating question pools, refreshing scenarios and checking results against how hires actually perform keep an assessment useful as AI tools improve.

Conclusion

The uncomfortable truth about Cluely detection is also the useful one: you cannot see an invisible overlay from a browser, so stop looking for a magic detector. Focus instead on what you control. Timed, hands-on, multi-step tasks shrink the read-back window that copilots depend on. Behavioral signals like timing anomalies and focus loss point you toward results worth a second look. Structured follow-up conversations and a clear human review process turn that evidence into fair decisions, without accusing anyone on a hunch.

If you want this approach without building it yourself, HireValid gives small businesses skills and cognitive tests built to resist AI assistance, with each candidate receiving a different, timed test served one question at a time. Its Integrity Score summarizes observed signals for a person to review, and nothing automated rejects a candidate. You can try HireValid free on three candidates, or compare monthly plans with no annual contract on the pricing page.

About the editorial team

Written and fact-checked by the HireValid editorial team. Sources checked . Read the editorial policy.

Does a low Integrity Score reject a candidate?+

No. Integrity signals may have innocent explanations, including connection issues or accessibility needs. A person should review the evidence and speak with the candidate before deciding.

What will candidates need?+

A browser and a reliable connection. Typing, spreadsheets and code tasks are intended for desktop. If an employer enables camera checks, candidates must receive a clear notice and a route to request an alternative.

Are the scores official qualifications?+

No. These are tools for hiring decisions. English results are CEFR-aligned level estimates, not official certificates. Cognitive scores are not clinical IQ results.

Keep exploring

LESS GUESSWORK. MORE GOOD PEOPLE.

AI can polish an answer.
Hire for the ability behind it.

Explore advanced hiring assessments and integrity tools built for the AI era.

Start freeOr try a sample test No credit card. No annual commitment.