Every résumé you receive now has been through a language model, and so has a large share of take-home submissions. Pretending otherwise is not a strategy.
Detection does not work and never will
Detection tools produce false positives at rates that would make you reject good candidates, and they will keep losing ground to generation quality. Building process on them is building on sand.
Interrogation works and always has
Ten minutes discussing a submission establishes authorship reliably. Ask why a decision was made, what was tried first, what would change under a different constraint. Authors answer; assemblers do not.
Résumés are now a filter, not a signal
Polished prose has stopped carrying information. Treat the résumé as a source of facts to verify — institution, batch, projects — and move quality assessment entirely into the conversation.
Live work beats submitted work
The reliability of asynchronous assessment has fallen sharply. A shorter live exercise now yields more signal per minute than a longer take-home, which reverses the trade-off of a few years ago.
Key takeaways
- Detection tools produce unacceptable false positives; do not build process on them.
- A ten-minute discussion of the submission establishes authorship reliably.
- Treat résumés as facts to verify, not as quality signals.
- Live exercises now out-yield take-homes per minute of assessment.
Questions
Should we ban AI use in assignments?+
Unenforceable, and it selects for candidates willing to say they did not use it. Assume use and design assessment that survives it.
Is AI-assisted work a red flag?+
Using tools well is a skill. Not understanding what the tool produced is the red flag, and interrogation surfaces that distinction immediately.