Both Sides Have AI Now. Verification, Not Detection, Is the Only Advantage Left.

By 2026, hiring has quietly become a conversation between two AI systems, with a candidate and a recruiter standing behind each one. Candidates use generative AI to draft, tailor and polish resumes and cover letters at a scale that would have taken a professional resume writer days per application. Employers use AI to parse, score and rank the results at a scale no human screening team could match. Trade press this year has covered this dynamic from both sides — recruiters describing application volumes that have outgrown their old screening workflows, and career-advice sites describing candidates whose AI-polished resumes get flagged, deprioritized, or waved through depending on which detector happens to be running that week. NACE, the National Association of Colleges and Employers, wrote about it under the heading “the ghostwritten candidate.” Vendor and staffing blogs have started calling it an “AI-on-AI arms race.”

We think that framing is right, and we think it’s missing the part that actually matters. An arms race implies the winning move is a better weapon: a sharper generator on one side, a sharper detector on the other. We don’t think that’s where this goes, and we built our own product on the opposite bet. Before we explain why, one editorial note, because it’s directly relevant to the topic: this piece deliberately does not repeat the specific percentages currently circulating about resume fraud and AI-detection rates. Several of the surveys behind those numbers are cited secondhand across dozens of near-identical blog posts, several of which read like they were themselves produced at scale for search visibility rather than reporting — which is its own small, uncomfortable illustration of the underlying problem: in a market flooded with AI-generated content, a statistic is only as trustworthy as the primary document behind it, and we were not able to trace several of this year’s most-quoted figures back to one. So we’re leaving them out rather than laundering them into a “cited statistic” that isn’t one. If you want to read the trade coverage yourself, NACE’s piece is a reasonable starting point.

Detection is losing, structurally

Here’s the part of the arms-race framing that we think is simply true, independent of any disputed number: text-based AI-content detectors are chasing a moving target by design. Every improvement to a detector trains, directly or indirectly, the next generation of tools built to evade it — the same adversarial dynamic that has played out in academic-integrity tooling for the last three years, and it does not resolve in the detector’s favor, because the generator only has to get past the check once per document while the detector has to be right every time. An applicant-tracking system that leans on “does this text look AI-written” as its filter is optimizing for a signal that degrades every time a language model improves, which in 2026 is roughly every quarter.

That’s the argument for not fighting this battle at the generation-detection layer at all. The more durable question isn’t “did AI write this sentence” — it’s “can this specific claim be traced back to something true.” That’s a different, more mechanical problem, and it’s one you can actually build a deterministic check for, on both sides of the hiring conversation.

What we built on the candidate side

NiceHire’s resume-tailoring feature uses AI to help a candidate adapt their resume to a specific job description — rewording a summary, reordering bullets to foreground relevant experience, tightening language. The generation step is the easy, unremarkable part; every resume tool does some version of it now. The part we think is worth writing about is what happens after generation, before the candidate ever sees a suggestion: every rewritten section runs through a separate, deterministic grounding check, implemented in plain code rather than asked of another model. It checks the rewrite against the candidate’s own existing profile — their actual work history, skills and prior wording — for anything that wasn’t there before: an invented number, a skill the profile never listed, a changed fact, an added bullet that describes a responsibility not otherwise evidenced, or language lifted straight from the job posting rather than the candidate’s own experience. If a rewrite fails that check, it is dropped outright and the original wording is restored — the candidate sees a plain notice that a suggestion couldn’t be verified against their profile, rather than a polished sentence they’d have to independently fact-check against their own resume. The check is exercised by a dedicated test suite covering both the rewrites that should pass and, deliberately, the ones that shouldn’t: invented numbers, invented skills, changed facts, added bullets, invented language claims, and job-description language laundered into the candidate’s own voice.

The reason it’s a separate deterministic check rather than “ask the model to grade its own work” is not a style preference. It reflects a finding that’s specific enough to cite properly: Huang et al.’s ICLR 2024 paper “Large Language Models Cannot Self-Correct Reasoning Yet” found that a model re-reading and revising its own output does not reliably recover ground truth it never had in the first place. A generator asked “is this true?” about its own generation is not a verification step; it’s the same process asking itself the same question twice. That finding is also, not coincidentally, why NiceHire’s own compliance process for public claims about our product works the same way — assertions are checked against artifacts in a file a human can read and a script can grep, not against another model’s opinion of them. We hold our resume-tailoring output to the same standard we hold our own marketing copy to.

What we built on the employer side

The same logic runs the other direction in NiceHire’s screening pipeline, though the shape of the check is different because the shape of the problem is different. Three things are true of every AI screening interview run through NiceHire, and each is enforced in the service code rather than described only in a policy: every applicant to a given role is asked from the same defined, ordered question set — the AI interviewer may ask follow-ups and rephrase a question on request, but the underlying question set and order don’t vary candidate to candidate. A completed screening stores a structured evaluation against the application: technical, communication and cultural-fit scores out of 100, an overall recommendation, and the specific strengths and gaps the evaluation found — not just a pass/fail verdict a recruiter has to trust blind. And when a candidate scores below a stage’s passing threshold, the default behavior is to route them to pending review for a human decision, not to auto-reject them; auto-reject exists, but it’s opt-in per stage rather than the out-of-the-box behavior.

None of that is a claim that our scoring is unbiased, or more accurate than a human recruiter, or an “audit trail” — we’re deliberately not making any of those claims, and we’ve had to correct our own marketing copy in the past for overreaching in exactly this direction. What it is a claim about is mechanism: what data a recruiter gets to see and check, and where a human decision sits by default in the pipeline. That’s a narrower claim than “fair AI hiring,” and it’s also one we can actually stand behind, because it’s checked against the code rather than asserted about an outcome nobody measured.

The actual asymmetry worth building for

Put the two sides together and the pattern is the same shape: don’t ask the generator to police itself, and don’t ask a black-box score to be trusted on its own. Build a second, separate, mechanical check — grounded in something outside the model that produced the claim — and give the human on the other end of the transaction the artifact that check produces, not just a verdict. On the candidate side, that artifact is “here’s what we couldn’t verify, and here’s your original wording instead.” On the employer side, it’s “here’s the specific evaluation and the specific gaps, and here’s why this candidate wasn’t auto-rejected.”

That’s a smaller, more boring answer than “smarter AI will win the arms race,” on either side of it. We think it’s also the only one that doesn’t degrade the next time someone ships a better model — because it was never a bet on out-writing or out-detecting the other side to begin with.

— NiceHire Team

Further reading: NACE, “The Ghostwritten Candidate: AI Fraud, Auto-apply, and the Fight for Authentic Career Readiness”.