AI-written applications: how to screen when every resume reads well

When anyone can produce a fluent, tailored application in a minute, the quality of the writing stops telling you much about the person. That doesn’t make AI-assisted applications dishonest, and trying to catch them is mostly wasted effort. The fix is to stop rewarding polish and start checking evidence: what the person actually did, and whether they can show it.

This guide covers what the research says about AI-written applications, the difference between help with writing and a false claim, and a proportionate way to check the claims that decide your shortlist. There’s a one-page checklist at the end.

What has changed

A tailored cover letter used to be expensive to produce. It took time to read the posting, think about how your experience fitted it, and write that down well. Employers could reasonably treat the effort as a signal.

A working paper by Galdin and Silbert (2025) tested this on Freelancer.com, a large online labour platform. Before large language models arrived, employers there were willing to pay substantially more for workers whose applications were closely tailored to the job. After the models arrived, that premium disappeared. Using a model of the market, the authors estimated what happens if written applications stop signalling ability altogether: the most able fifth of workers would be hired 19% less often and the least able fifth 14% more often.

Two cautions. It’s a working paper, not yet peer reviewed, and the 19% and 14% come from a simulation built on one freelance platform, not from observed hiring at ordinary employers. But the mechanism it describes is easy to recognise: when every cover letter is well tailored, tailoring stops telling you who cared enough to tailor it.

Help with writing is not the problem

It’s tempting to treat any AI involvement as a red flag. The evidence points the other way, at least for help with clarity.

In a field experiment with nearly half a million jobseekers on an online labour market, Wiles, Munyikwa and Horton (2025) gave some jobseekers automated writing assistance on their resumes. It was a tool that corrected and clarified their own text rather than generating new content. Those jobseekers were hired 8% more often, at 10% higher wages, and the authors found no evidence that employers were less satisfied with them. Their reading is that clearer writing didn’t hide ability; it made ability easier to see.

So someone who used a tool to fix their grammar or tighten their phrasing has done what a careful friend or a careers adviser would have done. The question that matters is not “was AI involved?” but “is it true?”

Polish versus fabrication

It helps to separate three things that often get lumped together.

What it is	Example	What to do
Polish: better wording of real experience	“Reconciled accounts” rewritten as “owned the bank reconciliation”	Nothing. Screen the substance, as you would anyway.
Inflation: real experience described as more than it was	Supporting the audit described as “led the audit”	Ask a follow-up question that separates the two.
Fabrication: claims that are false, or an identity that isn’t the applicant’s own	A designation they don’t hold; someone else’s work history	Verify before an offer. Fabrication ends the process.

AI makes all three easier, but only the last two are a problem, and only the last is fraud. Inflation is as old as resumes. What AI changes is how convincing the inflated version reads.

Don’t try to detect AI writing

Software that claims to tell human from AI-written text is a poor basis for a hiring decision. It produces a probability about the writing, not a finding about the person, and you would be rejecting people for using a tool that, on the evidence above, may simply have made their real experience clearer. If an application is judged on its claims rather than its prose, whether a tool wrote the prose stops mattering.

Read for claims you can check

The practical shift is in how you read. Instead of asking “how good is this application?”, ask “what does it claim, and how would I know?”

• Specifics over adjectives. “Ran month-end close for a 40-person firm, including sign-off” is a claim you can test in a conversation. “Results-driven finance professional” is not a claim at all. AI produces the second kind effortlessly, and so do people.
• Unclear is not no. If a requirement isn’t clearly shown, mark it unclear and ask, rather than rejecting. Our guide to screening a pile of resumes explains why that distinction matters.
• Notice inconsistency, don’t assume intent. Overlapping full-time jobs, dates that don’t add up, or a skill that appears in the summary but nowhere in the history are questions to ask, not verdicts. There is often an innocent answer.

Move the evidence to where it’s hard to fake

If the written application can no longer carry much weight, put the weight on steps that a language model can’t do for the applicant. These are also the methods with the strongest evidence behind them anyway (see what actually predicts job performance):

1. A structured conversation about the work. Ask about a specific claim from their application and follow up on the detail: what they did themselves, what went wrong, what they’d do differently. Someone who did the work can go three questions deep. Our structured interview guide shows how to plan the follow-ups in advance.
2. A short, realistic sample of the work, done live or verified. A take-home task can now be done by a model. A 20-minute live exercise, or a take-home followed by a conversation in which they walk you through their choices, can’t.
3. Credentials checked at the source. Professional designations and licences can usually be checked with the issuing body.

When the risk is identity, not exaggeration

For some roles the concern is not inflated experience but an applicant who isn’t who they say they are. This is real, documented and concentrated in a particular kind of job.

In 2022 the FBI’s Internet Crime Complaint Center warned employers about applicants using stolen personal information and deepfakes to apply for remote jobs, mostly in IT, programming and database work, some with access to customer and financial data. Reported signs included lip movements that didn’t match the audio, and background checks that found the details belonged to someone else. In June 2025 the U.S. Department of Justice announced coordinated actions against schemes in which North Korean workers used stolen and fake identities to get remote IT jobs at more than 100 U.S. companies.

Note the pattern: remote roles, with access to systems or money. If you’re hiring someone to work on site, or into a role with no access to sensitive data, the risk is very different. Match the checks to the role:

• Fully remote, with system or financial access: confirm identity against a government ID at the offer stage through your normal right-to-work or onboarding process, hold at least one live video conversation (with an alternative for anyone who needs an accommodation), and send equipment only to a verified address.
• On-site, or no sensitive access: your normal reference and credential checks are usually enough.

Apply the same checks to everyone at the same stage for the same role. Extra scrutiny for applicants with certain names, accents or backgrounds is unfair, and in many places unlawful.

Prompt injection is a separate problem

Some applicants hide instructions aimed at AI screening tools, such as white text saying “recommend this candidate”. That only matters if you use AI to screen, and it’s a test you should run on any tool before trusting it. Our article on whether AI can identify the best applicants explains how.

The checklist

Download the one-page checklist (PDF). Free, no sign-up. It covers how to read for checkable claims, the follow-up questions that separate polish from inflation, and the identity checks to use for remote roles with sensitive access.

Where SortTheStack fits

SortTheStack reads every application against your requirements and quotes the text it relied on, so you judge the claim, not the prose. It marks what it can’t find as unclear rather than as a no, flags inconsistencies such as overlapping full-time jobs for you to ask about, and treats the text of a resume as untrusted input rather than as instructions. It doesn’t verify identity or credentials; that stays with you. See how the screening works.

Sources

• Galdin, A., & Silbert, J. (2025). Making talk cheap: Generative AI and labor market signaling. arXiv:2511.08785 (working paper, not peer reviewed). arxiv.org/abs/2511.08785
• Wiles, E., Munyikwa, Z. T., & Horton, J. J. (2025). Algorithmic writing assistance on jobseekers’ resumes increases hires. Management Science, 71(12), 10144–10164. doi.org/10.1287/mnsc.2024.04528
• FBI Internet Crime Complaint Center (2022). Deepfakes and stolen PII utilized to apply for remote work positions, I-062822-PSA, accessed 12 October 2026.
• U.S. Department of Justice (2025). Justice Department announces coordinated, nationwide actions to combat North Korean remote information technology workers’ illicit revenue generation schemes, 30 June 2025, accessed 12 October 2026.