Your Applicant Tracking System Is Full of AI Resumes

Google paused part of its open source bug bounty programme last week, saying it had seen a significant rise in automated submissions, the vast majority of which were not valid. A company with Google's review capacity closed an intake channel rather than try to read its way through. Any business that posted a job in the last six months has been running a smaller version of the same experiment.

What the effort was doing

Nobody designed effort as a filter, which is why losing it is disorienting. A candidate who rewrote their covering letter for your role and mirrored the language in your posting had demonstrated something real: they read the posting, they wanted this specific job, and they were willing to spend time on it. None of that was stated anywhere in the document. All of it was inferred from the fact that the document existed.

Keyword matching in an applicant tracking system worked for the same reason. Matching your terms required reading your posting. The match was a proxy for attention, and the proxy held for twenty years because producing it any other way was slower than doing it properly.

Paste a job posting into a model today and you get a tailored application that mirrors every phrase, in under two minutes, with no reading required. The proxy now measures access to a tool. Everyone has the tool.

What changed and what to do

Every replacement in the third column has one property in common. It requires the candidate to do something at a specific moment, in response to something you chose, that cannot be prepared in advance in bulk. Effort has to re-enter the process somewhere, and after the application is the practical place left for it.

Detection is the wrong lever

The instinct is to buy a detector. The tools are unreliable enough that acting on their output is a legal exposure rather than a screening improvement, and the false positives are not randomly distributed. They fall on people writing in a second language and on candidates using accessibility tools, which turns a screening step into a defensible-practices problem. The broader bias and compliance picture around automated screening is in AI resume screening and hiring risk.

There is also no version of this where detection wins. The generated document gets better every quarter and the detector does not. Spending on the arms race buys a filter with a shelf life, and the money is better spent on an assessment step that does not care how the application was drafted.

Say how you expect AI to be used

A stated expectation in the posting does more work than a prohibition. Something close to: use whatever tools you like to draft and polish, the claims need to be yours, and we will go through the specifics in conversation. That is honest, it is enforceable at the interview rather than at the inbox, and it stops filtering for candidates who disadvantage themselves by following rules.

It also sets the tone for a job where the same tools will be in use. A candidate who used AI well to apply has demonstrated something relevant, and treating that as cheating sends an odd signal about how the work will be done.

Where AI belongs on your side

Using a model to summarise four hundred applications is reasonable and limited. It can extract dates, confirm whether a stated requirement appears, cluster candidates by background and flag the applications that answered your screening question with substance. Treat that as sorting rather than deciding, and keep a record of what it filtered out so the filter can be audited.

The review discipline matters more here than in most places, because a model reading uniformly fluent applications has no surface cues to work from, which is the problem described in quality assurance when AI writes the first draft. Sample what it rejected. A hiring funnel is one of the few workflows where the cost of a false negative is invisible and permanent.

A funnel that works at volume

The shape that holds up has three steps and no long form. The posting asks one specific question that requires looking at your business, answered in a few sentences. Candidates who answer it with substance get a short task with a real constraint, a few hours at most, with AI use declared. The task output is the basis for a conversation about how they did it.

That funnel is more work to design and considerably less work to run than reading four hundred polished documents. It also assesses the thing you are hiring for, which a covering letter did only by accident. If the underlying question is whether to hire at all, the comparison is in AI versus hiring an admin assistant, and the knock-on effect on junior roles is in hiring seniors and skipping juniors.

Frequently Asked Questions