Candidates using ChatGPT on hiring assessments
By Sergio Gualda · 5 min read · Updated
Some candidates will use ChatGPT on your assessment, and how much it matters depends on the format. A test with a correct answer or a long written response gives a model plenty to work with. A question about what this person would do, with no right option, gives it far less. Fix the format first, because heavy monitoring costs you candidates and only partly solves the problem.
How big the problem is
Fear runs ahead of the evidence. A 2026 study by the assessment company Clevry found that 62% of those surveyed believe candidates use AI to cheat, while 26% say they've seen evidence of it. Clevry sells assessments, so read that with some care, but the gap is telling.
The risk also varies by format. A take-home case study or a cover letter is exposed, because a model can produce a competent draft in seconds. An hour-long live conversation is far less so. Before you buy anything, work out which of your stages a model could complete by itself.
What each format is exposed to
Ordered from most to least exposed.
- Written take-homes and case studies. A model can draft them, and you can't tell how much of the result is the candidate's.
- Knowledge tests with a key. If there's a right answer, a model will often find it, which is why vendors pair these with proctoring.
- Coding tests. Models are strong here. Many teams now allow AI and judge what the person does with it.
- Situational questions with no right answer. A model can offer an opinion, but it doesn't know what this particular person would do, so there's less to gain.
- Live conversations with follow-ups. The hardest to fake, and the most expensive for you.
Why surveillance is a poor first move
Webcam proctoring, screen recording and locked browsers do deter some cheating. They also add a review burden, they can't see a second device, and they tell the candidate how you see them before you've met. The strongest people, who have other options, are the first to walk away.
They make sense for regulated exams or large-volume screening. For a product team hiring one designer, they're usually a bad trade.
What to do instead
Four changes that cost little.
- Ask for decisions and reasons about a specific situation, and avoid deliverables a model can produce.
- State your rule on AI in the invitation. Clear rules deter people, and they're fair to those who'd otherwise wonder.
- Follow up in the interview. Ask about the choice they made in the assessment, and ask what would change their mind.
- If the job involves using AI, let them use it and look at how they check its output.
How uxerhub handles it
uxerhub is about fifteen minutes of situations with no right answers, read against the team you describe. There's nothing to write or polish, nobody is recorded and nothing is locked down. If a result can't be read with confidence, we hold it back and return the credit. Every report ends with questions for the interview, which is where anything still uncertain gets settled.
Frequently asked
- Can you detect ChatGPT in a hiring assessment?
- Not reliably. Text detectors make mistakes in both directions, and accusing a candidate on that basis is risky. Designing the format well and following up in a conversation works better than trying to catch people.
- Should candidates be allowed to use AI in assessments?
- If the job uses it, often yes, as long as you say so and judge how they check the result. If you want to see their own thinking, say that instead. Either way, put the rule in the invitation.
- Do candidates use ChatGPT during live interviews too?
- Some try it on video calls. Follow-up questions about their own past work are the best defence, because a good answer depends on details a model doesn't have.