Take home assignments used to be a decent signal. Now a candidate can paste the prompt into an AI tool and get a polished answer in minutes. AI detectors aren't reliable either, and accusing a strong candidate by mistake is worse than missing a weak one.
The better question isn't "did they use AI?" It's "can they do the work, and do they understand it?"
Signals that don't work
- AI text detectors. They produce false positives and are easy to get around.
- Polish. A clean answer used to mean effort. Now it doesn't mean much either way.
- Banning AI. You can't enforce it, and many jobs expect people to use it well.
Signals that do work
- The process, not just the result. How did the answer come together? Steady work with edits and dead ends looks very different from one big paste.
- Follow up questions. Ask them to change one assumption, explain a tradeoff, or extend the solution live. Understanding shows quickly.
- Specificity to your context. Generic answers fit any company. Real thinking engages with your product, your data, your constraints.
Redesign the assignment
- Make it shorter and timed. 30 to 60 minutes on a focused problem beats a weekend project.
- Use your real context. A messy dataset from your domain, a bug from your codebase, a pricing decision you actually faced.
- Allow AI, and ask how they used it. Using AI well is a skill. Ask what they asked it, what they kept and what they threw away.
- Always follow up live. Ten minutes of discussion tells you more than the submission.
Watch the work, not just the answer
The strongest signal is seeing how someone gets to an answer. That's why Crossway records Arena solves as replays, with paste and tab switch tracking, and shows them to companies alongside the result. You see the thinking, the edits and the timing, which is much harder to fake than a finished file.
More on this in real work interview practice.
Questions
Can you detect AI written take home assignments?
Not reliably from the text alone. Process data, like a replay of how the answer was written, and live follow up questions are far better signals.
Should I ban AI in take home tests?
Usually not. It's hard to enforce, and using AI well is part of many jobs. Ask candidates how they used it instead.
What makes a better take home assignment?
Short, timed, based on your real context, and always followed by a live discussion.