How to Detect AI Cheating in a Software Engineer Take-Home Assignment

A take-home assignment shows AI-cheating signals when the submitted repository arrives as one dense commit instead of an iterative build history, the code style sits well above the candidate's stated experience level, and the solution matches what a general-purpose LLM produces for that exact prompt — verbatim variable names, exhaustive inline comments, and textbook error handling a self-taught junior wouldn't reach for on a first pass. Because take-homes are unsupervised and asynchronous, the browser session during authorship — not just the final diff — is where the real evidence lives.

Threat Model, Tells, and Evidence to Capture

Threat Model	Observable Tell	Evidence to Capture
Full solution generated by an AI coding assistant and submitted with light editing	Single large commit, or two or three commits, with no incremental "getting it working" history	Git commit timeline granularity; diff size per commit
Copy-paste from an AI chat tab into the IDE	Large paste events with no matching typing burst beforehand	Browser tab activity log; clipboard paste-volume telemetry
Style or idiom mismatch with the candidate's live-round code	Comments in a different voice or register; patterns the candidate can't explain in the debrief	Side-by-side code-style comparison across rounds; live debrief follow-up questions
Assignment completed in a fraction of the expected time	Submission timestamp shows total elapsed time far below the median for the task's stated complexity	Time-on-task tracking from assignment open to submit
Solution matches a public reference answer for a well-known assignment	High textual or structural similarity to a public repo or blog post solving the identical problem	AI-content and plagiarism similarity scoring against public sources

Interviewer Script

Use this as a live follow-up after submission, before advancing the candidate to the next round:

"Walk me through the commit where you first got the tests passing — what was broken before that?"
"You used a memoization decorator here — why that pattern over a simple dict cache, and what's the tradeoff?"
"If I asked you to add support for negative numbers right now, where would you start?"
Silence, or generic textbook-sounding answers to any of the above — especially paired with a completion time far below what the task warrants — is the strongest combined signal of AI authorship.

FAQs

Isn't using Copilot normal for engineers day-to-day? Yes, and that's exactly why the debrief question matters more than a blanket ban: the goal is verifying the candidate can explain and extend what was submitted, not policing which tools touched the keyboard.

What if the candidate has genuinely fast baseline skill? Fast, skilled candidates still show incremental commit history and can defend every design choice instantly. AI-submitted work tends to stall on "why," not "what."

Can commit history or file metadata be faked? It can be edited after the fact, but browser-session telemetry captured during authorship — tab switches, paste events, time-on-task — is far harder to fabricate retroactively.

How common is this? Fabric's dataset of 19,368 interviews (July 2025–January 2026) flagged 38.5% overall for AI-cheating signals, rising to 48% in software engineering specifically. Take-homes are a major contributor because they're unsupervised.

Should we drop take-home assignments entirely? Most teams keep them for signal on real-world coding style, but pair them with instrumented authorship tracking and a mandatory live defense of the submission.

What's the fastest way to add this without slowing down the process? A five-minute scored debrief on the flagged submission, using the script above, is usually enough to separate genuine authors from AI-only submissions without adding a new interview round.

Related Guides

Get the Evidence Before You Extend an Offer

Unsupervised take-home assignments are exactly where AI-generated submissions hide best — there's no interviewer watching, so the browser session itself has to do the work. Neuroxa Browser Proctoring captures tab history, paste events, and time-on-task during the assignment window, then generates an authorship-confidence score your team can review before the debrief.