Fifty-four percent of talent leaders now say AI-generated resumes, applications, and interview prep have made it harder to assess a candidate's true capability — and they say it in the same breath that more than 90% of them report running AI across sourcing and screening, while fewer than 5% call the results transformational (ManpowerGroup, 2026). Read those three numbers together and the story is not "AI is helping us hire." It is: you spent two years automating one side of the funnel, your candidates spent the same two years automating the other, and the net effect on the thing that actually matters — reading who can do the job — went backwards.
That is the trap most mid-market hiring operations are sitting in right now. The instinct when screening feels broken is to buy a better applicant tracking system or bolt on another AI filter. The ManpowerGroup and Everest Group data says that instinct is aimed at the wrong layer. This is not a tooling gap. It is a signal problem, and no amount of screening automation fixes a signal that has been made fakeable at scale.
Both Sides of the Funnel Automated at Once
Start with the symmetry, because it is the part almost nobody prices in. When you deployed AI to source, rank, and screen, you assumed the inputs on the other side stayed constant — that a resume still carried roughly the same information it did in 2022. It doesn't. The candidate now has the same class of tool you do. Generative AI writes the resume, tailors the cover letter to your posting, drafts answers to your screening questions, and rehearses the interview.
The ManpowerGroup/Everest Group study — a survey of 80 C-suite, CHRO, and senior talent-acquisition leaders across the US and UK — puts a number on the consequence: 54% report that AI-assisted candidate behavior is making it harder to accurately assess true capability (ManpowerGroup, 2026). That is a majority of senior hiring leaders telling you the input has degraded. Both sides automated at once, and the arms race netted out where arms races usually do — a lot more spend on both sides, and no lasting advantage to either.
The Signal Didn't Get Noisier — It Got Fakeable
There is a critical distinction here that changes what you do about it. The problem is not that there is more noise in the funnel. It is that the resume-and-application layer is now AI-inflatable — and once a signal can be manufactured cheaply, ranking candidates on it stops measuring the thing you care about.
Think about what your screening AI actually optimizes. It scores and ranks candidates on the strength of their written application: the keywords, the phrasing, the apparent match to the job description. When that written layer was produced by the candidate's own effort, it was a weak but real proxy for diligence and fit. When it is produced by a model in nine seconds, the ranking increasingly rewards prompt fluency — how well the candidate (or their tool) reverse-engineered your posting — rather than job fit. You are still getting a crisp, confident ranked list. It is just measuring the wrong variable. That is the most dangerous failure mode in selection: not a system that visibly breaks, but one that keeps producing confident output after the thing it measures has quietly become meaningless.
This is also why "fewer than 5% transformational" is not a maturity problem that another quarter of adoption solves (ManpowerGroup, 2026). You cannot extract a transformational result from a layer whose signal has been hollowed out. More screening horsepower applied to a fakeable input yields faster, more confident sorting of noise.
The Volume Problem Hiding Underneath
The signal problem rides on top of a volume problem, and the two compound. Because applying is now nearly free, application counts have detached from genuine interest. One recruiter reported a business-analyst role that would historically have drawn 100 to 150 applications pulling in more than 2,000 this time — a 13-to-20-fold jump for a single opening (The Irish Times, 2026). That is not a surge of qualified interest. It is the mechanical output of candidates running the same automation you are.
The cost lands on your team's time. In the same reporting, a Greenhouse survey found 38% of recruiters now spend half their effort filtering out junk, spam, and completely unqualified applications (The Irish Times, 2026). Stack that on the signal problem and you get the full picture: more applications than ever, each one a weaker indicator than ever, and a recruiting team burning half its capacity separating machine-generated volume from human intent. Buying more screening AI to process that faster does not shrink the problem. It industrializes it.
Why a Better ATS Won't Fix This
The tempting response — a smarter ATS, a sharper resume-parsing model, an AI that "detects AI" — loses the arms race by design. Every detector you deploy is a target the generation tools optimize against, and the generation tools iterate faster and are used by millions of candidates. You would be committing your budget to defending the exact layer that has already been compromised.
The deeper reason it fails is that it accepts a broken premise: that the resume is where capability lives. It never fully did — and now it demonstrably doesn't. The fix is not to read the fakeable layer more cleverly. It is to stop weighting the fakeable layer so heavily and move decision weight onto inputs a candidate cannot cheaply manufacture at scale. That is a design change in how you screen, not a purchase.
There is also a quieter cost to staying in the arms race: false rejections. When your ranking rewards prompt fluency, the candidate who is genuinely strong but a mediocre self-marketer — or who declined to run their application through an optimizer — sinks below the one who did. You are not just failing to catch inflated applications; you are actively down-ranking real capability that didn't bother to game the layer. In a market where a single opening now draws thousands of applications, those false negatives are invisible and unrecoverable. You never learn which qualified person your fakeable signal buried.
Re-Weight Toward Inputs Candidates Can't Fake at Scale
Here the research is unusually settled, and it predates the AI panic by decades. The reference point is Schmidt and Hunter's meta-analysis synthesizing roughly 85 years of selection research, which found that work sample tests, structured interviews, and cognitive-ability measures are among the strongest predictors of job performance — while the resume-adjacent signals of education and years of experience are comparatively weak (Schmidt & Hunter, 1998). A structured, job-representative assessment carries far more predictive validity than the application document — and it was the better instrument even when the application document was honestly written.
AI shifts this from a best-practice argument to an urgent one. The methods that hold up are precisely the ones that are expensive to fake at scale. A model can generate a flawless resume in seconds; it cannot sit your actual candidate through a job-representative work sample and produce their genuine reasoning under structured conditions. A validated psychometric assessment measures traits and aptitudes that a polished application simply doesn't reveal. The signal you lost on the resume layer is recoverable — but only by moving the measurement to a place the automation can't reach.
Concretely, "inputs candidates can't fake at scale" means two things a mid-market ops team can actually deploy. First, a structured work sample: a short, job-representative task scored against a consistent rubric, so you are watching the person do a slice of the actual work rather than reading their claim about it. Second, a validated psychometric assessment: a measured read on the aptitudes and traits that predict performance in the role, administered under conditions the candidate can't outsource. Both move the decisive signal off the document and onto behavior — and both, notably, are more legally defensible than resume screening because they sample the job directly.
The One Move for This Quarter
You do not need to rip out your ATS or re-architect hiring to act on this. You need to change where the decision weight sits on one role.
So the concrete move is narrow and testable: pick your highest-volume open req — the one drowning in AI-inflated applications — and insert one job-representative work sample plus a validated assessment early in the process, then re-weight your ranking so those results outrank the resume score. Use the resume layer only to screen out the obviously unqualified; let the un-fakeable inputs decide who advances. Then instrument it against a comparable role you left on resume-first screening, and watch two numbers over a quarter: quality-of-hire signals like ramp speed and early performance, and the share of recruiter time reclaimed from junk-filtering.
If the re-weighted role produces better hires with less wasted screening effort, you will have proven, inside your own funnel, the thing the ManpowerGroup data is pointing at: the answer to a hollowed-out signal was never a better screener for the fakeable layer. It was to stop trusting the layer your candidates learned to automate — and to measure the one thing they still can't. The 54% who say they can't read capability anymore are right about the problem. They are looking for it in the one place AI guaranteed it would no longer be.
