It is cheating when the employer has said not to, and many never say. Anthropic bans AI on take homes unless it states otherwise; Rootly and Speak expect it. A field guide to 51 AI engineering hiring processes found that most take homes state no AI policy at all. When the brief is silent, ask in writing, and disclose what you used.
Which employers say no, and which say yes?
Anthropic's candidate guidance, last updated 10 July 2025, asks candidates to complete take homes without Claude unless Anthropic indicates otherwise. Deloitte UK's integrity guidance says its online assessments must be completed without AI assistance.
On the other side, Rootly tells candidates that AI use is expected and strongly encouraged on its take home, and asks for the AI session transcripts. Speak replaced its engineer evaluation with take home projects in which candidates build a light product spec using AI agents, according to LeadDev. Even Anthropic runs one performance engineering take home that explicitly allows AI.
Our company policies page lists each company's rule, with the source and the date we checked it.
What if the brief says nothing about AI?
That is the common case. Alexey Grigorev's field guide to AI engineering hiring, covering 51 companies with disclosed processes, found 17 that use a take home. Only one explicitly allowed AI tools, none explicitly banned them for take homes, and most did not mention a policy at all.
So ask. A one line email to the recruiter, asking whether AI tools are allowed and whether they want to know how you used them, settles it and leaves a record. If they allow it, say what you used in the README. If they do not answer, disclosure is the safer default.
Why have take homes become hard to grade?
Because the models got good at them. Anthropic's Tristan Hume wrote in January 2026 that by May 2025 more than half of the candidates on its performance take home would have scored better by handing the whole task to Claude Code, and that a later model, Claude Opus 4.5, matched the best human score within two hours. The test is now on its third version, built from unusual puzzles that simulate novel work.
Smaller companies hit the same wall. The Pragmatic Engineer reported in April 2025 that a startup hid an instruction in its take home brief that only an AI assistant would follow; all four completed submissions contained it, and three of those candidates said they had not used AI. That is one anecdote, but it shows why undisclosed use is risky.
How do employers who allow AI judge a take home?
They look past the code. Rootly asks for a repository, the AI transcripts and a video under five minutes, and its reviewers read the transcripts for how the candidate broke the problem down and corrected the AI. Speak looks at the directions a candidate takes on an open ended brief. The field guide reports that take homes in AI engineering roles are often followed by a live walkthrough of 45 to 90 minutes, where you defend your decisions.
In each case the code is necessary but not enough. What you are judged on is whether you can explain and stand behind it.
What counts as cheating even when AI is allowed?
Submitting work you cannot explain, hiding AI use when the employer asked you to disclose it, editing a transcript to look better, or having another person do the work. Simon Willison's rule for his own code is a good test: he will not commit code he could not explain to someone else.
Practise the honest version. Our practice rounds are take homes of 60 minutes with AI allowed, and each ends with a prompt that asks your own agent to review the session transcript, so you see what a reviewer like Rootly's would see.