Take-home tests are done by AI now. What do you replace them with?
Replace the take-home with a short task the candidate does while you can see how they work: 20 to 30 minutes of real work, with AI allowed. You stop guessing who wrote the answer and start looking at what the person did.
What went wrong with the take-home
A take-home test hands the candidate a brief and asks for a finished result. That result is all you see. Today a candidate can paste the brief into a chatbot and send back what it produces, polished, on time, with no sign of who did the thinking. Some will do exactly that. Many honest candidates will not, and they pay the price: they spend an evening on the task while others spend two minutes.
Banning AI does not fix this. You cannot enforce it on work done at home, and it tests a way of working your new hire will not use on the job.
What to ask for instead
The fix is not a harder take-home. It is a task where the process is visible. Four things to look for:
- Role skills. Can they do the core of the job with real material in front of them?
- AI collaboration. Do they use the assistant well, with clear questions and sensible use of the answers?
- Critical verification. Do they check what the AI tells them before they rely on it?
- Judgment under uncertainty. When the material is incomplete or contradicts itself, do they decide sensibly and say what they are unsure of?
You cannot read these from a finished document. You can read them from watching the work.
Take-home versus observed work sample
| Take-home test | Short observed work sample | |
|---|---|---|
| Where it happens | At home, unobserved | In a simulation, where the work is recorded |
| What you see | The finished result only | What the candidate did, step by step |
| AI use | Hidden or banned | Allowed, and part of what you look at |
| Candidate time | Often an evening | 20 to 30 minutes of real work |
| Effort for you to grade | A reviewer reads each submission | A report per candidate, ranked |
| Who finishes it | Candidates with spare time | Most candidates who are interested |
What the research says
The idea is not new. A 2022 re-analysis of decades of selection research, by Sackett, Zhang, Berry and Lievens in the Journal of Applied Psychology, placed work sample tests among the five strongest predictors of job performance. A take-home is a work sample with the observation taken out. Put the observation back and you keep what made it useful. We do not yet have a validity study of our own, and we do not claim one.
How WorkProbe does it
WorkProbe is a work-sample assessment. Paste in a job description and it generates a job simulation from it, or bring your own scenario. The candidate does 20 to 30 minutes of real work: messy documents, simulated colleagues, and an AI assistant they are free to use.
Some of the assistant's answers are wrong or incomplete on purpose. You see who checks and who forwards. WorkProbe reports what each candidate did, with the evidence, across the four skills above. You get a ranked shortlist in plain-word ratings, never scores or percentages, and interview questions anchored to what each person did. WorkProbe ranks and records. A person makes the hiring decision.
What to do this week
- Take the role where you use a take-home now.
- Replace the take-home with a short simulation for the same role.
- Say in the invitation that AI is allowed.
- Compare the shortlist with the one your take-home would have given you.