Recruiters have noticed the same shift: applications are more polished, more uniform, and more numerous. Cover letters read well and say little. Two candidates describe the same responsibilities in nearly identical phrasing.
The common response is to look for a way to detect and penalise it. That response fails on all three of its assumptions.
Detection does not work
AI-text detectors are unreliable in both directions, and their errors are not random. They flag non-native English speakers at higher rates, because the features they associate with generated text — regular sentence structure, conventional phrasing, limited idiom — also describe careful writing by someone working in a second language.
Using such a tool to screen candidates means introducing a systematic disadvantage against a group, on the basis of an unreliable signal, for a behaviour that is not misconduct. That is a discrimination exposure with no upside.
Penalising it is not defensible
Consider what candidates are actually doing. They are using an available tool to present their real experience more clearly, frequently after being told for years that their CV was the reason they were not getting responses.
Penalising that penalises candidates who lack the confidence, the network, or the professional coaching that better-resourced candidates have always had access to. AI assistance has, if anything, levelled a field that was tilted by access to help.
There is a real problem underneath, but it is not "the candidate used a tool". It is fabrication — claiming experience that does not exist. That problem predates AI, and the response to it has always been the same: verify claims rather than judge prose.
What actually changed
Three things, and only the third is a problem.
- Volume rose. Applying is cheaper, so people apply to more roles. This shows up as rising applications per hire and is a throughput problem, not a quality one.
- Writing quality stopped being a signal. It was always a weak proxy for capability in most roles, and it now carries close to no information. That is a loss of a signal you should not have been using.
- Specificity became the discriminator. Generated text is fluent and generic. It describes responsibilities well and outcomes poorly, because it does not have the outcomes.
What to screen on instead
Specific outcomes with numbers
"Responsible for improving the onboarding process" is generatable from a job title. "Cut onboarding from eleven days to four by removing two approval steps" is not, because it requires knowing what happened. Ask for it directly in the application.
One targeted question
A single short question tied to the specific role, answered in a few sentences. It can still be answered with assistance — but the assistance needs the candidate real material to work with, which is exactly the filter you wanted. See screening questions that filter.
A short work sample
The most reliable option where the role allows it. Assess the work rather than the description of the work. Keep it short — under an hour — and pay for anything longer.
Verification in the interview
Every claim on a CV should be discussable. Someone who genuinely cut onboarding from eleven days to four can describe the two approval steps, who objected, and what broke. Someone who cannot has told you what you needed to know, and it did not require a detector.
The stance worth taking publicly
Say it in the posting: we do not mind how you drafted your application, we assess specific evidence and we will discuss what you have written. This is honest, it removes anxiety for candidates who are unsure, and it sets the expectation that claims will be examined.
It is also more effective than any policy prohibiting AI use, which is unenforceable and mostly signals that the process has not adapted.
What this means for screening design
The direction of travel is away from evaluating documents and toward evaluating evidence and work. Applications are getting cheaper to produce, so the information they carry per page falls. Processes that lean on written self-description will get progressively worse signal; processes built on structured questions, work samples and evidence-based interviews will not.
That shift was worth making before any of this — see structured interview loops. AI-written applications have made it urgent rather than optional.