Most applications now contain AI-written content. What that does to screening.

For most of recruitment's history the CV was a reasonably honest artefact. Not a complete one, and not always accurate, but it carried real signal: how someone described their own work told you something about how they thought, and the effort involved in tailoring an application told you something about how much they wanted the job. Both of those signals have now largely evaporated, and the industry has not finished working out what replaces them.

The numbers are stark. Reporting this year puts AI-generated content in roughly 78% of job applications, with about a quarter of CVs recruiters see now identifiably AI-written. Only 37% of hiring managers still consider a CV a reliable indicator of talent. Meanwhile application volume per role has climbed sharply, because the marginal cost of applying has fallen to almost nothing.

That combination — more applications, less signal per application — is the defining screening problem of 2026. It is not a problem you solve by screening harder.

Why detection is the wrong instinct

The first response most agencies have is to try to identify AI-written applications and discount them. It is an understandable instinct and it fails for three separate reasons.

The first is that detection does not work reliably. AI-detection tools produce false positives at rates that would be unacceptable in any other part of a hiring process, and they systematically misfire on non-native English speakers and on anyone whose writing is simply formal. Rejecting a candidate on a probabilistic guess about their prose is not defensible, and in some jurisdictions it is not lawful.

The second is that it punishes the wrong behaviour. A candidate who uses a language model to clean up their CV is doing the 2026 equivalent of asking a friend who is good at writing to look it over. That has never been a disqualifier. The behaviour worth catching is fabrication — claimed experience that did not happen — and that is a different problem with a different solution.

The third is the most uncomfortable. Research this year found that large language models score CVs they generated themselves 67% to 82% higher than equivalent human-written ones. If your screening stack uses an LLM to rank applications, it has a measurable preference for LLM-written applications. Detection and ranking pull against each other inside the same system.

What the CV still tells you

The useful reframe is to stop treating the CV as evidence and start treating it as a claim. A claim is still worth having — it tells you what the candidate believes is relevant and how they are positioning themselves — but it is the beginning of an assessment rather than a substitute for one.

Read that way, some parts of a CV survive the AI era intact. Dates, employers, and job titles are checkable facts that a language model does not invent unprompted. Career shape — the trajectory across roles, the pattern of tenure, the direction of increasing responsibility — is hard to fabricate coherently and is not what generation tools change. Specific, unusual, verifiable detail is more credible than fluent generality, and always was.

What no longer carries weight is polish. Prose quality, keyword density, formatting, and the presence of well-phrased achievement bullets are now free. Any screening heuristic that rewards them is measuring tool access rather than capability.

Where the signal moved

If the document is weaker evidence, the assessment has to move somewhere. In practice it has moved to three places.

• Structured conversation. A fifteen-minute call with three specific, non-generic questions about work the candidate claims to have done will separate real experience from generated experience faster than any screening tool. This has always been true; it is now load-bearing rather than confirmatory.
• Demonstrated competency. Work samples, structured exercises, and technical assessments measure what someone can do rather than what they can describe. The shift to skills-based hiring started for reasons unrelated to AI, but the collapse in document signal has accelerated it considerably.
• Verified history. Reference checks, employment verification, and credential checks have quietly become more valuable, because they touch facts outside the candidate's control.

None of these are new. What is new is that they can no longer be treated as later-stage confirmation of a judgement already formed at the CV stage — because the judgement formed at the CV stage is now much less reliable.

The volume problem underneath

Even with better assessment, the arithmetic has to work. If a role attracts three hundred applications and each meaningful assessment takes fifteen minutes, no amount of methodological rigour makes that tractable. The signal problem and the volume problem have to be solved together.

This is the argument for matching that works on meaning rather than on text. Keyword and boolean filters were always a proxy for relevance, and in a world where candidates optimise their CVs against those exact filters, the proxy degrades badly — you end up surfacing the candidates best at keyword-matching rather than the candidates best suited to the role. Semantic matching compares the substance of described experience against the substance of the requirement, which is harder to game with phrasing alone.

The critical constraint is that whatever ranks the pile has to show its reasoning. A ranked list with no explanation asks a recruiter to trust a black box at precisely the moment when trust in automated evaluation is at its lowest. Placr's matching engine returns a dimension-by-dimension breakdown — which requirements are met, which are not, and what evidence sits behind each — so the recruiter is reviewing an argument rather than accepting a score. We wrote about why we built it that way in how we built AI matching that explains itself.

The candidate side of this

There is a fairness dimension that agencies should think about before their clients ask. Candidates are using these tools in large numbers, and they are also on the receiving end of them: recent survey work found only around 26% of candidates trust AI to evaluate them fairly, and a majority of rejected applicants report receiving no human feedback at all.

Notably, candidates are not asking for AI to be removed. Survey data suggests the most common preferences are for the same level of AI with more transparency, or more AI with stronger human oversight at the decision points. What they object to is opacity — being evaluated by a process nobody will describe, and rejected by one nobody will explain.

That is a solvable problem, and solving it is a competitive position rather than a compliance chore. An agency that can tell a candidate what was assessed, by what, and who made the final call has an answer that most of the market currently does not. It is also, increasingly, the answer regulators are moving towards.

What to change this quarter

Three practical changes cover most of the gap.

1. Stop scoring applications on prose quality and formatting, in both your human process and any automated ranking you run. Reward specificity and checkable detail instead.
2. Move first-stage assessment from document review to a short structured conversation or a work sample, and accept that this makes early-stage screening slower per candidate — the offsetting saving comes from getting the shortlist right.
3. Make sure whatever tool ranks your inbound pile explains its ranking, and that a human is accountable for the decision at every point where someone is rejected.

The agencies that come out of this well will not be the ones with the best AI detector. They will be the ones that noticed the document stopped being evidence and rebuilt their assessment around things that still are.