In this article
Before you start
The interview process explains the stages of a loop and who judges each one.
Why interviewers ask this
Candidates ask it because the stakes are real: a flagged answer can count against you, and the rules differ by company and by round. An interviewer, in turn, wants to know whether the reasoning on screen belongs to the person in the call.
A weak answer to the question says "yes, they always know" or "no, nobody can tell." Both are claims about every employer at once. A strong answer separates three things: what software records, what a person notices, and what the company's own rules allow. Some formats allow AI by design, as the page on the AI-assisted coding interview explains, so the first question is never "can they tell" but "what did they ask me to do."
Detection from first principles
- False flag
- A false flag is a warning raised about a person who did nothing wrong, so it points a human reviewer at an honest candidate.
In plain words
Think of a smoke alarm. It makes noise when it senses something that often goes with a fire, but burnt toast sets it off too. The analogy breaks in one place. A person walks into the kitchen to check a smoke alarm. A flag about an interview can count against you before anyone asks you about it. Everything below covers what raises the alarm and what a careful person does next.
In one precise sentence, interviewers can tell if you use AI only to the degree that software signals or live conversation produce evidence, and neither produces proof alone.
Step 1: Test platforms may record signals
Some online assessment platforms offer monitoring features that record how a timed test was taken. Read the instructions the employer sent you and the platform's own start page before the test, because they describe what applies to your assessment.
Input: your activity on the test page. Output: whatever record the platform keeps, if any. A record is not a verdict. You now know the first possible source of flags: a platform that writes down what happened during a timed test.
Step 2: Interviewers notice mismatches in conversation
In a live round, the strongest signal is the conversation itself. Interviewers ask why you chose an approach, what changes if an input grows, and what you would do if a constraint flipped. A candidate who can recite a polished answer but cannot adjust it under a new constraint looks the same whether the polish came from a book, a friend or a tool.
This is reasoning, not a published measurement: a strong follow-up shows whether the explanation is yours. The page on looking at notes during a video interview covers a related rule question that matters even before AI enters.
Step 3: A flag is a question for a person, not proof
A flag from software or an interviewer's hunch should lead to a human question, and the reason is a base rate. A base rate is the share of a group that truly has a trait before any test is applied, and it decides how much a positive result means.
Take an illustrative round: 1,000 candidates, 5% of whom use AI against the rules, a detector that flags 80% of them and wrongly flags 2% of everyone else.1 That flags 40 of the AI-using candidates and 19 honest candidates, so 32.2% of flagged people are honest.1 If the false-flag rate is 10%, the same detector flags 95 honest candidates beside those 40, and 70.4% of flagged people are honest.1
Real detectors do err against some writers: Liang et al. (2023) show that detectors of AI-written text are biased against non-native English writers. So careful reviewers treat a flag as a reason to ask more questions, and that is why the follow-up decides.
Step-by-step description
- A test platform may record signals.
- Interviewers notice mismatches.
- A person decides what a flag means.
- Rules decide what was allowed.
- Some test platforms offer monitoring features.
- Interviewers notice mismatches in conversation.
- A flag is a question for a person.
- Company rules decide what is allowed.
| Quantity (illustrative) | 2% false-flag rate | 10% false-flag rate |
|---|---|---|
| AI users flagged | 40 | 40 |
| Honest candidates flagged | 19 | 95 |
| Share of flagged people who are honest | 32.2% | 70.4% |
Text description of the diagram
- In the illustrative round 40 AI users and 19 honest candidates are flagged at a 2% false-flag rate.
- At a 10% false-flag rate 95 honest candidates are flagged.
All three inputs are made up to show the arithmetic; they are not measured rates for any product or company.1
Step 4: Company rules decide what is allowed
Whether AI use is a problem depends on the instruction you were given, not on whether anyone noticed. Some formats ask for no outside help, others invite an assistant. The Anthropic interview process page and the guide to applicant tracking systems and AI show how differently stages can be treated. When the written instructions are unclear, ask the recruiter before the round and keep the reply.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →The Staff-level answer
Can interviewers tell if you use AI during a round?
Sometimes, and not through any one tool. On an online assessment some platforms offer monitoring, and in a live round the interviewer tests whether my reasoning is mine by changing a constraint and asking why. A flag from software is evidence, not proof, because detectors have false positives, so a careful reviewer follows up with a person. What matters more is the rule: I read the instructions, ask the recruiter if they are unclear, and prepare so that I can explain every line or decision I give, whatever help I used while studying.
What makes it Staff-level: it separates signal from proof, names the false-positive problem, and ends on rules and understanding instead of evasion.
The rubric
| Dimension | Weak answer | Strong answer | Staff answer |
|---|---|---|---|
| Correctness | "They always know." | Names signals and says none is proof. | Adds base rates and false flags. |
| Depth | Software only. | Software and live follow-ups. | Explains why the follow-up is the real test. |
| Trade-offs | Ignores policy. | Notes rules differ by company. | Says to ask the recruiter and keep the reply. |
| Communication | Defensive. | Calm and specific. | Ends on being able to explain every step. |
Three variants of this question
Can an AI detector prove I used AI?
No. A detector gives a score, and honest writers are sometimes flagged; Liang et al. (2023) report bias against non-native English writers. A score should start a conversation, not end one.
What if the company allows AI?
When the company allows AI, the grading usually moves to how you direct, check and explain the work. Follow the format described in the AI-assisted coding interview guide and say aloud what you accept and what you reject.
How do interviewers test whether an answer is really yours?
Interviewers test it by changing the problem: a larger input, a new constraint, a failure case, or a request to justify one design choice. A strong candidate adapts without restarting. The interview skills page covers how to show that reasoning clearly.
Common mistakes
- "They can never tell, so the rules do not matter." (wrong) Rules decide consequences even when nothing is noticed. (right) "I follow the written instructions and ask when they are unclear."
- "A detector score proves it." (wrong) Scores have false positives, as the base-rate table shows. (right) "A flag is a reason to ask me more questions."
- "I will memorise a polished answer." (wrong) A follow-up with a changed constraint exposes it. (right) "I learn the reasoning so I can adapt the answer."
- "I will not mention that I prepared with AI tools." (wrong) Hiding ordinary preparation adds stress with no benefit. (right) "If asked, I say how I prepared and what I can explain myself."
Follow-up questions
Can interviewers tell if you use AI on an online assessment?
Sometimes. Some platforms offer monitoring, and a reviewer may then look at what was recorded. Check the instructions you were given, and treat any flag as a question a person will ask you.
Can interviewers tell if you used AI on a take-home assignment?
A take-home is usually discussed afterwards, so be ready to explain every part of your submission and how you would change it; see the take-home assignment example.
Should I tell the interviewer I used AI to prepare?
Using a tool to study is ordinary preparation, and you can say so if asked. What counts in the round is that you can explain every decision without help.
What if I get a follow-up I cannot answer?
Say what you know, state the assumption you would check, and work the new case aloud. A visible reasoning process reads better than a polished guess.
Where do I check the rules for my round?
Start with the recruiter's email and the platform's start page, then read the company's own page, such as the Anthropic interview process, and ask when something is unclear.
- Interviewers can sometimes tell, through platform monitoring where it is used and through follow-up questions in live rounds.
- No single tool proves AI use; flags have false positives, and a person should review them.
- Base rates matter: a rare behaviour plus a modest false-flag rate puts many honest people among the flagged.
- Company rules decide what is allowed, so read the instructions and ask the recruiter.
- Prepare so you can explain every step of your answer when the constraint changes.
Sources
- GPT detectors are biased against non-native English writers (arXiv 2304.02819, v3) — arXiv (Liang et al.), 10 Jul 2023
- GPT detectors are biased against non-native English writers (Patterns 4(7), 100779) — Patterns, Cell Press (Liang et al.), 10 Jul 2023
