Take-home tests when candidates use ChatGPT or Claude.
Candidates can do most take-homes with AI, whatever the rules say. So decide what you want to learn: say that AI is expected, make the task real work, and ask for the conversation with the AI as well as the result.
What companies do with take-homes now that candidates use AI
There is no common rule yet. Companies that hire for AI work have gone in opposite directions.
- Ban AI Anthropic asks candidates to do take-homes “without Claude unless we indicate otherwise”. anthropic.com, last updated 10 Jul 2025
- Expect AI Canva expects engineering candidates to use AI tools such as Copilot, Cursor and Claude in technical interviews: “We want to see how well candidates collaborate with AI to solve problems.” canva.dev, 11 Jun 2025
- Say nothing Many take-home instructions do not mention AI at all, so candidates have to guess what is allowed.
Should you ban AI on take-home tests, or allow it?
Four ways to handle AI on a take-home, and what each one tells you about the candidate.
| What you learn | The problem | Right when | |
|---|---|---|---|
| Ban AI | How they work without AI, if they follow the rule. | It is hard to check. Honest candidates do it by hand; others do not, and you cannot tell which is which. | The job is done without AI, or you test a skill AI must not touch. |
| Say nothing | Not much. Some use AI and some do not, and you compare the two. | Candidates guess the rule. The careful ones guess “no”. | Never. Say something. |
| Allow AI, collect only the result | A finished piece of work, which AI makes look good. | You cannot see who did the thinking, or whether they noticed when the AI was wrong. | The result is hard to get right even with AI. |
| Expect AI, collect the conversation | How they briefed the AI, what they checked and what they changed. | Takes longer to read. Someone has to read and score it. | The job is done with AI. |
Can you tell if a candidate used ChatGPT on a take-home?
Not reliably, and not well enough to turn someone down for it. OpenAI withdrew its own AI text classifier in July 2023 “due to its low rate of accuracy” (OpenAI). Code and plans are harder still to judge than prose.
Instead of detecting AI
Allow it, and ask for the conversation. Then the question is not whether they used AI, but how: what they gave it, what they checked, and what they changed.
Instead of a proctored take-home
Ask about one moment in the conversation in the interview. Someone else doing the work is hard to hide when you ask why they trusted a number at turn 12.
How to design a take-home test when candidates use AI: six steps
- 1Say what the AI rules are, in the first line.“Use any AI tool you like. We expect you to.” Or the opposite. Either way, nobody has to guess.
- 2Use a real task, with real files.A bug in a real codebase, a budget with last year’s numbers, 400 invoices and six contracts. A puzzle with a known answer is one AI can just solve.
- 3Build in one place where AI tends to go wrong.A number that has to match a source, a definition that changes the answer, a detail buried in a long document. That is where you see who checks.
- 4Keep it to one to three hours, with a deadline of days.Not a strict timer. Say how long it should take, and score nothing that rewards working longer.
- 5Ask for the AI conversation, and a short note.The conversation with the AI, plus a few lines on what they checked and what they did not. Say up front that you will read it.
- 6Spend the interview on one moment.Pick a turn in the conversation and ask about it. It is hard to answer for work you did not do.
What candidates should hand in with an AI-allowed take-home
In Lemonsignal, candidates do the task in Claude Code, Codex, Claude Cowork or ChatGPT and the conversation is recorded for you, as a session replay with the moments that matter marked and six skills scored.
Questions about take-home tests and AI
Should candidates be allowed to use ChatGPT or Claude on a take-home test?
If they will use AI in the job, yes, and say so in the brief. A take-home that bans AI is hard to enforce and tests a way of working they will not use once hired. Then ask for their conversation with the AI, so you see how they worked, not only the result.
Can you tell if a candidate used ChatGPT on a take-home?
Not reliably from the finished work. AI text detectors are not accurate enough to judge a candidate on: OpenAI withdrew its own AI text classifier in July 2023 because of its low rate of accuracy. It is easier to allow AI and ask for the conversation.
Are take-home tests still useful now that AI exists?
Yes, if they change. A take-home that AI can finish alone shows little. One built on real work, with AI expected and the conversation collected, shows how the candidate works with AI, which is what the job now asks for.
How long should an AI-allowed take-home take?
One to three hours of work, with a deadline of a few days rather than a timer. AI makes candidates faster, so a task that took a day before may now take a couple of hours.
What should candidates hand in with an AI-allowed take-home?
The work, their conversation with the AI (most AI tools can share or export one), and a short note on what they checked and what they did not. In Lemonsignal the conversation is recorded for you as a session replay.
Is a live interview better than a take-home when candidates use AI?
Neither is better; they show different things. Live shows how someone works under time pressure as it happens. A take-home shows how they work when nobody is watching, on a task closer to the real job.