When Every Candidate Uses AI, What Is a Resume Still Worth?

By Victor Hoang · Updated Oct 10, 2026

A resume used to tell you two things at once: what a person had done, and whether they could be bothered to say it well. The second signal is gone. Any candidate can now produce a tailored, fluent application in under a minute, so fluency no longer separates anyone from anyone.

The answer isn't to screen documents harder. Detection is a losing game, and the better the tools get, the less a clean document proves. The answer is to move decision weight toward evidence that costs real effort to fake: a work sample, a structured live evaluation, and references you actually verify. Treat the resume as an index of claims to test, not as proof.

Key Facts

• Applications per job opening rose from 116 in 2022 to 244 in 2025, an increase of 111%, across more than 6,000 companies and 640 million applications in Greenhouse data. That is one hiring platform's customer base, not the whole labor market.
• In the same data, recruiters per organization fell from 10.43 to 4.62 over 2022 to 2025. More applications are landing on fewer reviewers.
• After Freelancer.com introduced a one-click AI proposal tool in April 2023, employers' willingness to pay for highly tailored proposals fell sharply, and the signal no longer predicted job success among hired workers. The setting was short freelance digital jobs, not full-time roles.
• A separate study of the same kind of tool found the correlation between cover-letter alignment and callbacks dropped by 51% after launch, with employers leaning more on prior work history.
• In the 2022 revision of the selection-validity research, structured interviews score 0.42 against 0.19 for unstructured interviews and 0.07 for years of job experience as predictors of job performance.

The Signal Was Always Effort

Economists describe a good signal as one that's costly for weak candidates and cheap for strong ones. A thoughtful cover letter worked that way because writing one took time. If you got sixty applications in an afternoon, the ones that engaged with your actual problem stood out, and they usually came from people who cared.

AI removed the cost. The Freelancer.com research is the cleanest evidence of what happens next. Before the platform's AI writing tool, employers paid more for workers who sent customized proposals, because those proposals predicted successful completion. Afterward, that link broke. The authors, Anaïs Galdin and Jesse Silbert, then modelled a world where everyone can write an expert application. In that simulation, top-ability workers were hired 19% less often and bottom-ability workers 14% more often. That last part is a simulation, not an observed outcome, and it comes from a marketplace for small online jobs. But the mechanism travels. When a signal becomes free, the market stops rewarding the people who used to earn it.

Look at what employers did in response. In the second study, the correlation between how well a letter matched the posting and whether the applicant got a callback fell by 51%, and employers shifted toward other signals such as prior work history. That's rational. But the risk is easy to see: if prose stops helping, reviewers may lean harder on employer and school names, which are cruder filters and can narrow who gets seen.

Volume Makes It Worse

The pressure isn't only about signal quality. Greenhouse's benchmark data shows applications per job climbing from 116 to 244 between 2022 and 2025, while recruiter headcount per organization dropped by more than half. We can't say AI caused all of that, and the source doesn't claim to isolate it. But a team reading twice the applications with half the people will skim, and skimming a pile of equally polished documents is close to random.

So the instinct to add an AI screener on top is understandable. We take that question up separately in should AI screen job candidates. The short version: a model that ranks documents is ranking the very thing that just lost its meaning.

What the Selection Research Says Instead

Industrial-organizational psychology has spent decades asking which hiring methods predict later job performance. The landmark compilation was Schmidt and Hunter's 1998 review. In 2022, Sackett, Zhang, Berry and Lievens revisited it and concluded that earlier estimates had been inflated by overcorrection for restriction of range. Most procedures kept their rank, but validity estimates fell by roughly .10 to .20, and structured interviews came out on top.

Their comparison table is the most useful single page for this decision:

Three caveats keep this honest. First, these are averages across many studies and jobs, and the structured interview's spread is wide: the paper reports a lower bound of .18 on its 80% credibility interval, so a badly designed one can disappoint. Second, the research measured predictors of job performance and doesn't test the AI-era question directly. It predates widespread LLM use, so how these methods hold up against AI-assisted candidates is still an open question. Third, the authors themselves note that the strongest predictors, which they list as structured interviews, job knowledge tests, empirically keyed biodata and work samples, are all job-specific measures. That's the useful pattern: the closer the assessment sits to the actual work, the better it predicts.

Notice what's missing from the table. There's no row for "polished resume" or "cover letter". Neither is a validated predictor in this literature, and that was true before AI. What changed is that hiring teams could previously get away with using documents as a proxy for effort. Now they can't.

A Framework: Claims, Samples, Structure, Verification

Here's a mental model for redesigning a funnel. Call it the four-layer evidence stack. Each layer asks a different question, and each is progressively harder for a language model to fake on the candidate's behalf.

Two of these layers need a closer look.

Work samples. Keep them short, tied to a real task, and scored against a rubric written before you see any candidate. An illustrative example: for a revenue operations analyst, hand over a messy anonymized pipeline export and ask for three findings and one recommendation in 60 minutes. The point isn't that candidates can't use AI. They probably can and will. The point is that you see how they frame the problem, what they decide to trust, and whether they can defend the answer out loud afterwards. That last step is the filter. A candidate who pasted a prompt and submitted the output can't explain why the second finding matters.

Verified references. Be straight about the evidence here: the 2022 paper lists reference checks among the procedures it excluded for insufficient information, so we can't point to a validity coefficient. Their value is different. A reference is a human who can confirm a specific claim, such as "led the migration, 40 people, shipped on time", that no language model can supply on a candidate's behalf. Ask about particular events, not general character, and treat references as a check against what you saw in the work sample and interview rather than as a standalone score.

Where This Goes Wrong

Moving weight is not the same as piling on process. Three failure modes show up.

Work samples that are really free labor. A four-hour unpaid project asked of fifty candidates is a reputation cost and screens out people with caregiving or second jobs. Keep it to an hour or two, pay for anything longer, and only give it to a short list.

Structured interviews that drift back to chat. The research advantage comes from consistent questions, anchored rating scales and independent scoring. If each interviewer improvises, you've rebuilt the unstructured interview, the 0.19 line in the table above, with extra paperwork.

Treating AI use itself as the offence. Many candidates use AI the way they use spellcheck. If your own team uses it daily, penalizing the same behavior in applicants is incoherent. The more useful question, which we explore in should candidates use AI in interviews, is what judgment the candidate adds on top of it. And that links to a bigger shift in what AI-native new hires expect and what you can reasonably expect from them.

What to Do Next

Different executives own different parts of this.

CHRO or head of talent

• Audit your current funnel. For your last ten hires, record which stage actually changed the decision. If it was the resume or the cover letter, you're exposed.
• Move the resume to a stage-one claims list. Give hiring managers a one-page template that turns resume claims into three probing questions.
• Standardize interviews for your ten highest-volume roles: fixed questions, a 1 to 5 anchored scale per question, and scores submitted before the debrief.

COO or hiring-function owner

• Build one work sample per high-volume role, with a rubric. Pilot it on the next five hires and compare scores with 90-day manager ratings.
• Put a cap on time per candidate and publish it. Candidates can tell when a process respects them.
• Decide in writing what AI assistance is allowed in the sample, and ask candidates to say what they used.

CEO

• Ask for one number from the people team each quarter: of the candidates who reached final interview, what share had completed a work sample? Make it rise.
• Ask whether reviewers are leaning on school and employer names to cope with volume. If so, the pile of polished documents may be narrowing your talent pool rather than widening it.
• Pair this with the skills picture in the AI skills gap executives are getting wrong and skills-based talent strategy, since a skills-first funnel only works if you define the skills.

None of this requires new software. It requires deciding that the document isn't the decision.

Learn More

• Should AI Screen Job Candidates?
• Should Candidates Use AI in Interviews?
• AI-Native New Hire Expectations
• The AI Skills Gap Executives Are Getting Wrong
• Skills-Based Talent Strategy
• The Broken Bottom Rung