A piece from Peterson Technology Partners has been going around this week, and it names the dynamic more precisely than most: candidates are using AI to get through screening, companies are using AI to do the screening, and the escalation between the two is making hiring worse for both sides rather than better.

The numbers in it are worth sitting with. In HackerRank's survey data, 81% of US developers say it's hard or very hard to find a job despite the number of open roles. 76% say AI has made technical assessments easier to fool. 78% say the assessments in use don't measure real work. Greenhouse CEO Daniel Chait puts it about as plainly as it can be put: "We've got this tragic situation where each side has a problem. They're using AI to solve their own problem, but in ways that make the problem worse."

I want to pick up where that piece leaves off, because the obvious conclusion — build a better detector — is the one I think is wrong.

Detection is a race you lose slowly

Every AI-writing detector that works becomes the test the next generation of writing tools is tuned to pass. That isn't a hypothesis. It's the history of spam filtering, plagiarism checking, and ad blocking, compressed into about eighteen months. The detector improves, the generator improves against it, and the durable result is that both sides spend more to stay in the same place.

Hiring has a worse version of this problem than spam does, because the cost of a false positive isn't an email in a junk folder. It's a qualified person who never hears back and never finds out why.

The escalation has a second cost that's easier to miss. As both sides automate, the people lose their read on what's happening. A recruiter looking at a stack of uniformly polished applications can't tell which polish is competence and which is tooling. A candidate who gets filtered can't tell whether it was the wrong keyword, the wrong background, or nothing at all. Two systems are negotiating, and neither person can see the negotiation.

What actually broke

The instinct is to say the problem is that candidates are using AI. I don't think that's it. Someone using a tool to write a clearer account of real work is doing what everyone with a tool does.

The problem is that screening went opaque at the same time it went fast. When a score comes out of a model that can't say why, there's nothing for either side to check. The recruiter can't verify it, so they either trust it completely or ignore it entirely. The candidate can't contest it, so they optimize for whatever they guess it wants — which is precisely the behavior that starts the arms race.

Opacity is what makes the escalation rational. If the reasoning is visible, gaming it stops being the winning move, because the gap between a real qualification and a well-written claim becomes something you can see rather than something you have to infer.

What showing the work looks like

This is the part of our own product I'd defend hardest, so discount it accordingly — but it's also the most concrete thing I can offer.

• The criteria are written down first. Required, preferred, bonus, with weighting you set. The score isn't a verdict from nowhere; it's an answer to a question you asked in words.
• Eligibility and knockout rules run before the score, and they're rules a person wrote. Work authorization, licensure, location — things that are genuinely binary get handled as binary instead of being smuggled into a number.
• Every score cites what it's built on. You can open it, read the evidence it pulled from the application, and disagree with it. Disagreeing with it is a normal thing to do, and the interface expects you to.
• Authenticity signals flag a candidate for a human to look at. They don't reject anyone. Most of them aren't reading the resume prose, because the prose is the cheapest part to polish.

The common thread is that there's no point in the flow where the model makes the call. Advancing or rejecting a candidate is a physical action a recruiter takes. We don't offer auto-reject below a threshold or auto-advance above one, and that's a decision we made rather than a feature we haven't gotten to.

What this doesn't fix

I don't want to oversell it. Explainable scoring does not detect an AI-written resume, and I'd be skeptical of anyone who claims theirs does reliably.

What it changes is what you're evaluating. If a score has to point at specific evidence — shipped work, a described project, a verifiable credential — then a fluent application with nothing behind it scores like what it is. Polish stops working as a proxy for substance, because the system was never grading polish in the first place.

It also doesn't make the volume problem disappear. Nobody can read every application, and pretending otherwise helps no one. What a score can honestly do is set the reading order and explain itself well enough that a recruiter can tell when the order is wrong.

If you're evaluating tools this quarter

Three questions. The answers tend to be revealing.

1. Can you show me why this candidate scored what they scored, in terms I could repeat back to them? If the answer is a confidence interval, that's not an answer.
2. What happens automatically to a low score? Anything other than "nothing" means the vendor has made a decision about your process on your behalf.
3. What is the flag actually reading? If the whole system runs on resume text, it's grading the thing that is easiest to generate.

The arms race that piece describes is real, and I don't think it ends because one side ships a better model. It ends in the places where a person can see enough to make the call themselves.

The original piece: AI Resumes Meet AI Recruiters: Dispatches from the Automated Hiring Cycle, Peterson Technology Partners, September 9, 2026.