You prevent AI cheating in interviews by asking questions an assistant can't answer for the candidate. interviewing.io ran the experiment: candidates used ChatGPT secretly in technical interviews, and not a single interviewer mentioned concerns about cheating. Pass rates on verbatim LeetCode questions hit 73%. On custom questions, 25%.
That second number is the whole strategy. Here's how I'd design an interview loop in 2026 so that AI help stops paying off.
Key Takeaways
- Standard puzzle questions are solved by any assistant. Custom, context-heavy questions are not, and the interviewing.io data shows the gap.
- Follow-ups beat first answers. Ask "why," then "what would break," then "what did you try first."
- Decide your AI policy and say it out loud. Some teams allow AI and test judgment; others ban it. Ambiguity is what cheaters exploit.
- Put at least one moment in the process that's hard to fake: a live conversation about the candidate's own work, on camera.
Why the old interview broke
Tools like Interview Coder openly market themselves as "100% Invisible to Screen-Recording" and list HackerRank and CoderPad among the platforms they target. Its creator recorded himself using it in an Amazon internship interview, NBC News reported. Whether every claim holds up doesn't matter much. The assistant only needs to work some of the time.
Big companies reacted. Sundar Pichai said on the Lex Fridman podcast that Google would "introduce at least one round of in-person interviews." If you hire remote engineers across borders, flying everyone in isn't an option. You need a remote loop that holds up.
How to prevent AI cheating in interviews
1. Retire public questions. If a question is on LeetCode, the assistant has seen it. Write problems from your own codebase: a simplified version of a real bug, a real schema with a real performance problem.
2. Ask about their work, not the textbook. "Tell me about the last production incident you handled. What did the logs show first?" An assistant can produce a plausible incident. It can't handle the fifth follow-up about it.
3. Follow up three levels deep. First answer: what. Second: why that and not the alternative. Third: what would you change now. Cheaters are fluent on the first and collapse on the third.
4. Change the constraints mid-answer. "Now assume the table has 400 million rows and you can't add an index." A relayed answer lags behind the new constraint.
5. Make them explain code out loud. Reading code someone else wrote and explaining what it does, line by line, is slow to fake in real time.
6. Say the rules. If AI is allowed, say which tools and watch how they use them. If it's banned, say so in writing before the interview. We wrote about the allowed version in how to vet developers who use AI coding tools.
| Question type | Easy to cheat with AI? | Better replacement |
|---|---|---|
| Known algorithm puzzle | Yes | Custom problem from your domain |
| "Explain X concept" | Yes | "Where did X bite you in production?" |
| Take-home with no follow-up | Yes | Take-home plus a live walkthrough of their own code |
| System design, generic | Partly | Design with your real constraints, changed mid-answer |
| Their own past project | Hard | Same, with three levels of follow-up |
Where we draw the line at Ruzora
Every engineer on our bench takes a graded coding challenge and an AI interview answered on camera, out loud, in the engineer's own words. The two are separate evaluations with separate scores, and an engineer needs to pass both to reach our bench.
Then the client interviews the engineer, after the service agreement, with full identity. That interview is where steps 2 to 5 above belong. A vendor's process is one layer. Yours is another.
A Concrete Version
A 20-person dev-tools startup redesigns its loop for a senior Go role after a bad hire who was fluent in interviews and lost in week two.
- Old loop (3 hours): 2 LeetCode mediums, a generic system design, a culture chat.
- New loop (3 hours):
- 45 min: a bug from their own repo, simplified. The candidate can use docs, not assistants, stated in the invite.
- 60 min: system design of their actual ingestion pipeline, with the constraint changed twice.
- 45 min: walkthrough of the candidate's last project, three follow-ups per answer.
- 30 min: candidate reads and explains a 60-line function from the codebase.
Same interview time. Say the next six candidates include two who would have cleared the old loop. In the new one they show up in the walkthrough: excellent first answers, empty third follow-ups.
The Honest Counterpoint
Custom questions cost real time to write and calibrate, maybe a day per role, and they leak once enough candidates see them. Long walkthrough interviews also favor confident talkers; a quiet, strong engineer can look worse than a fluent, average one.
And "no AI" is getting harder to justify. Your engineers use assistants every day. Testing them without one measures something they'll never do at work. The better question for many teams is whether the candidate understands what the assistant produced, which is exactly what follow-ups test.
Frequently Asked Questions
How do you prevent AI cheating in interviews?
Use custom questions from your own domain, ask three levels of follow-up, change constraints mid-answer, have candidates explain code out loud, and state your AI policy in writing before the interview.
Can interviewers tell when a candidate uses ChatGPT?
Often not. In interviewing.io's experiment with 32 audio-only interviews, no interviewer raised concerns about cheating. Custom questions cut the cheaters' pass rate to 25%, which is why question design matters more than spotting.
Should we just allow AI in interviews?
Some teams do, and test how the candidate uses it. Either policy works if you say it clearly. See interview cheating detection tools if you want software support on top.
The Bottom Line
Preventing AI cheating is mostly interview design: custom problems, deep follow-ups, and a live conversation about the candidate's own work. If you'd rather start from engineers who passed a graded coding challenge and an on-camera AI interview, see our engineers or request a shortlist.
Roberto Espinoza is CEO of Ruzora, which helps US startups hire pre-vetted senior LATAM engineers, with a vetted shortlist in 72 hours. See available engineers.
