You cannot detect your way out of AI-assisted candidates

A new category of hiring tool has appeared in the last year: software that promises to detect AI-assisted or fully synthetic candidates. Flag the polished-but-shallow resume. Spot the scripted answer. Notice the unnatural pause. Catch the candidate using a model in real time. The pitch is intuitive, and the underlying problem is real. But the strategy has a structural flaw that no amount of model quality can fix.

Detection is an arms race, and it is one the detector loses by default. Every signal a detector learns to flag is a signal the next generation of tools learns to smooth over. Pauses get shorter. Resumes get more specific. Answers get more natural. The detector is always reacting to last month's tells, while the tools it is trying to catch improve every week. You are not buying a solution. You are buying a position in a race with no finish line.

The false positives are the real cost

Step back from the cat-and-mouse framing and a worse problem appears. Detection tools do not output truth. They output probability, and they apply it to people. A candidate who writes clearly, prepares thoroughly, and is naturally concise looks, to a detector, exactly like a candidate using AI. A non-native English speaker who drafts answers carefully looks suspicious. A neurodivergent candidate whose cadence is unusual looks suspicious. The tool cannot tell diligence from deception, because the surface features are identical.

When the output is a flag on a person, every false positive is a qualified candidate quietly pushed out of the funnel. You will never see the cost, because rejected candidates do not file reports. The damage is invisible and one-directional, which is exactly why it is dangerous.

The question detection answers is the wrong question

Detection asks: did this candidate use AI? That is the wrong question, for two reasons. First, it is increasingly unanswerable. Second, and more importantly, it does not matter as much as the industry assumes. The thing you actually want to know is not whether a candidate used a model to polish a CV. It is whether they can reason through the work they claim to have done.

Those are completely different questions, and only the second one predicts whether someone will be good at the job. A candidate who used AI to tighten their resume but genuinely led the migration they describe is a strong hire. A candidate who wrote every word themselves but cannot explain a single trade-off in work they claim to have owned is not. Detection cannot tell these two apart. It is measuring the wrong variable.

What beats detection

There is a different strategy, and it does not require winning any arms race. Instead of trying to catch the fake, build an interview that the fake cannot pass and the real candidate can. The mechanism is simple: anchor every question to a specific claim on the candidate's own CV, then test the trajectory behind it.

AI is excellent at producing a correct-looking answer to a generic question. It is weak at reconstructing the specific, messy path of having actually built something: the dead end that was abandoned, the constraint that changed halfway through, the trade-off that was rejected, the thing that broke in production at 2am. You cannot pre-generate that, because it is anchored to a real history that either exists or does not.

1. 1Extract the concrete, testable claims from the candidate's CV in their own words.
2. 2Generate a probe for each claim that targets the reasoning, not the recall.
3. 3Add a perturbation: change a constraint partway through and watch whether they re-reason or recite.
4. 4Let a human interviewer mark which evidence signals appeared. The decision stays human.

Notice what this approach never has to do. It never has to decide whether the candidate used AI. It never has to flag a person. It never has to surveil behaviour or measure timing. It simply tests whether the lived experience is there. A candidate who did the work walks through it easily. A candidate who did not, with or without AI, runs out of texture fast. The interview does the filtering, not a probability score.

Not an AI detector. Not a fraud score. A better technical screen.

The detection vendors are right about one thing: the old static-question interview is broken in a world of cheap, polished answers. They are wrong about the fix. The fix is not to catch the candidate. It is to ask better questions, anchored to what each candidate actually claims, and to keep a human in the seat that decides. That is a position you can hold indefinitely, because it does not depend on staying one step ahead of anyone.