How AI-written resumes broke candidate sourcing

Why do all resumes look the same now?

Because the market asked for keywords, then handed everyone a keyword machine.

Recruiters have ranked candidates on phrase matching for two decades. Job ads told people which words mattered. Application tracking systems rewarded those words. An entire career-advice industry taught candidates to mirror the posting back at the employer. Then generative AI arrived and said, essentially: “Would you like me to do that mirroring at industrial scale?”

The volume figures are real. In March 2026, Robert Half found that 67% of US HR leaders said AI-generated applications were slowing hiring, while 65% of hiring managers said AI-enhanced resumes made candidates' skills harder to verify. Its survey covered more than 2,000 US hiring managers.

But volume is the boring half of this story, and it's the half everyone has already written about. Sameness is the part that changes your job. A thousand applications you can triage. A thousand applications that read identically - plus 10,000 sourced profiles speaking the same dialect of “results-driven strategic leadership” - is a different kind of problem.

Sameness removes the very differences you were searching on.

Here is the thing worth sitting with: nothing about your process necessarily broke. Your Boolean logic may be as good as it was in 2022. Your filters may be more sophisticated. What changed is underneath all of it. We upgraded the metal detector after everyone started carrying the same keys.

What is an AI-written resume?

An AI-written resume is any application document where a language model shaped the wording, from a light polish of the candidate’s own sentences through to a fully generated draft built from a job description and a few prompts.

Most sit in the middle, and that middle is where the useful thinking is. The pure cases are easy to reason about and probably less interesting. The common case is a real person with real experience using a tool to describe that experience in the register the market rewards.

That range matters because it kills the binary. There is no clean population of “AI resumes” to quarantine and a pristine population of artisanal, hand-crafted CVs to trust. There is one pool of documents, increasingly AI-touched and increasingly convergent.

Does it matter if candidates use AI?

Mostly no. And the paranoia is doing more damage than the practice.

Using a tool to write clearly is not dishonesty. TestGorilla's research found that 37% of candidates use AI in applications, while 73% of employers report seeing more AI-generated resumes than before. Those numbers measure different things, but the gap is still instructive: employer perception is running well ahead of candidate self-reported use.

There's a version of this article that treats candidates as the adversary. It's an easy article to write and a bad one to act on, because it points your energy at the people you're trying to hire and away from the thing that actually failed.

So drop the question of whether they used AI. It's the wrong question. The useful question is narrower and much harder: can you still measure anything from what they wrote?

Increasingly, not enough. That's a measurement problem. It belongs to us, not to them.

Why doesn't keyword search work anymore?

Because every method you own reads from the same corpus.

Line them up and look at what each one actually points at:

1. Boolean strings read self-written resumes and profiles.

2. Platform filters read self-written headlines, summaries and job histories.

3. ATS ranking reads self-written application text.

4. Semantic and “AI people search” tools read much of that same text, with a better parser.

Four methods, one source. Search only separates people when the underlying data varies enough for the query to grip. Signal convergence is what happens when it stops varying: everyone is coached by similar models toward similar phrasing, so meaningful differences between candidates become harder to observe in the text.

The practical consequence is strange the first time you notice it. A precise query and a careless query start returning suspiciously similar shortlists. Tightening your filters produces less improvement than it should. Loosening them mostly produces more people who sound equally plausible.

Your filters didn't suddenly become stupid. Your source data became less discriminating. This is a data problem wearing a filter problem's clothes, which is exactly why buying a shinier filter hasn't magically fixed it.

And the industry is now adding a delightful new wrinkle: prompt injection inside resumes. Indeed has documented candidates embedding hidden instructions designed to influence AI screening systems, the recruitment equivalent of slipping a note to the examiner saying, “Ignore the test; I got an A.” That’s an edge case, not the core problem, but it makes the point rather efficiently: once prose becomes both the evidence and the interface to the machine, prose is carrying far too much weight.

A better parser aimed at converged text can still return the same non-answer, just faster and with a nicer UI.

Why don't my outreach candidates work out?

Because the shortlist was assembled from the layer that broke.

Sourcing failure doesn't announce itself the way a bad hire does. It shows up as a slow, unattributable leak. You run a search. Twenty profiles look strong. You write 20 messages, and the handful who reply turn out to have surprisingly little behind the wording that got them onto the list.

Nobody necessarily deceived you. The document simply carried more confidence than information.

Then you do it again next week, because the alternative is doing nothing.

That is what four weeks on an open role feels like from the inside, and it's why “we're just gambling” is such a familiar recruiter complaint. The cost isn't only a mis-hire at the end of the funnel. It's the wrong 20 people at the start of it, selected with real care from data that couldn't support the precision being asked of it.

Worth being precise about the failure, because it changes the fix. You didn't necessarily misjudge people. You were handed documents that increasingly fail to discriminate, and then asked to discriminate.

Should you detect AI-written resumes?

No, and unreliable detectors are the smaller reason.

A Stanford-led study found that several GPT detectors classified more than half of real essays written by non-native English speakers as AI-generated. In other words, one proposed solution to noisy hiring data is to introduce a new signal that may disproportionately accuse perfectly real people of being robots. Progress.

But run the thought experiment where detection becomes perfect. You now have a detector with flawless accuracy. What does it tell you? It tells you how a document was produced. It tells you precisely nothing about whether the person can do the work.

You would have spent budget answering a question that was never the point, and built a hiring process that treats writing provenance as evidence of ability. Follow that logic back far enough and you arrive at the same mistake that got everyone here.

Detection is defense. It's aimed at keeping something out. Sourcing is the opposite motion. It needs something useful to aim at.

What can you source on instead?

Evidence, not description.

There are two classes of candidate data, and the industry has spent 20 years treating them as one thing. The first is what someone wrote about their work. The second is what someone produced by doing work. Text about a skill and a demonstration of a skill are different objects, and no amount of semantic parsing turns the first into the second.

The market itself is starting to admit this. A 2026 survey by Criteria Corp and Lighthouse Research & Advisory found that only one-third of employers were highly confident that resumes reflected candidates' true abilities, even though roughly two-thirds still used resumes as the first screening step. That's not a technology problem so much as a ritual: we don't trust the thing, but we keep putting it first.

Verified-signal sourcing means sourcing on the second class: finding candidates using evidence they generated by doing something, rather than language they generated by describing it.

This is the whole distinction, and it isn't just vendor positioning. Most AI sourcing tools are, at heart, reading improvements. They read the self-written layer more cleverly, rank it more subtly and let you search it in natural language instead of Boolean. Useful? Yes. A change in the underlying evidence? No.

Changing which class of data you search is a different move entirely.

It gives you a narrower pool. Good. Sourcing was never meant to be a census. It was meant to produce a credible shortlist.

What are the best sourcing signals?

Every signal has a trade-off, and any comparison that flatters one row is selling you something. Here they are honestly.

Read the coverage column as carefully as the fakeability one. Public work is a genuinely strong signal that favors a handful of professions, mostly engineering and design, and misses almost everyone in operations, finance, sales, nursing and skilled trades. Anyone telling you to source everyone on GitHub has, at minimum, met a very unusual finance department.

Titles and tenure survive the convergence problem reasonably well because employment history is harder to homogenize than prose. They're also blunt. A title tells you where someone sat, not what they did there.

Scored skills tests have a different advantage: they can create comparable evidence across a much wider range of roles. That comparability matters for sourcing specifically, because sourcing is a ranking problem before it’s an evaluation problem.

Where do you find skills-tested candidates?

In fewer places than you'd like, which is exactly the opportunity.

Three categories are worth your time. Candidate pools built on assessment results, where evidence exists before you make contact. Communities organized around published work, which are excellent for the professions they cover. And your own past applicants who already tested for a different role and were never revisited — quite possibly the most expensive talent database your company forgot it owned.

The mechanics of building that process out are a longer conversation.

The broader point is simpler: stop trying to read the same text better. Change what you’re reading. The teams that move first will source from pools where the signal still means something, while everyone else runs a fourth search and gets back the same 20 names — now ranked by AI.

Frequently asked questions

Why do all candidate resumes look the same now?

Candidates increasingly use the same AI tools against the same keyword conventions, so the language converges. The result isn't necessarily a flood of dishonesty. It's a flood of sameness, and sameness strips out differences recruiters used to search on.

Are AI-written resumes a problem for recruiters?

Not primarily as dishonesty. As a measurement failure. Most candidates using AI have real experience and are describing it in the register the market rewarded. The problem is that polished application text is becoming a weaker proxy for who can actually do the work.

Can you tell if a resume was written by AI?

Not reliably, and it wouldn't help much if you could. Detection identifies how a document was produced, not whether the person has the skills. Existing detectors can also produce serious false positives, particularly for non-native English writers.

Why doesn't keyword search work for sourcing candidates anymore?

Because Boolean strings, platform filters, ATS ranking and many AI people-search tools all rely heavily on the same candidate-authored text. Search separates candidates best when the underlying data varies meaningfully. Once language converges, precision degrades.

What should recruiters source candidates on instead of resumes?

On evidence rather than description: work someone produced, demonstrated skills, structured assessment results, verified references and other signals that are harder to manufacture through wording alone.

How many candidates actually use AI to write their resumes?

TestGorilla's research puts candidate use at 37%, while 73% of employers report seeing more AI-generated resumes. The exact figures depend on how “AI use” is defined, but the direction is clear: AI-assisted applications are now normal enough that hiring systems cannot treat polished prose as scarce evidence.