Nobody Sent Out the New Rulebook

Somewhere in the last two years, hiring turned into a closed loop that nobody designed on purpose. A candidate opens a model and asks it to tailor a resume to a job description. A recruiter, buried under the resulting volume, opens a different model and asks it to screen the resumes down to a shortlist. A hiring manager looks at the score the second model produced and treats it as a fact about the candidate. Three people, three tools, and at no point in that chain does anyone actually look at a person. I don’t think any single actor in this chain made a bad decision. I think the chain itself has quietly stopped being about hiring at all.

Start with the candidate side, because that’s where the volume begins. Applying to a role used to cost something, an evening, a real attempt to understand the company and say something specific. AI collapsed that cost to nearly zero. A candidate can now generate a plausibly tailored resume and cover letter for a hundred roles in the time it used to take to write one, and in a market where the number of people worth hiring shrinks slower than the number of applications, that is a rational thing to do. But the trade is not free. What a hand-written application used to signal, that someone cared enough about this specific role to spend real time on it, is gone. Volume replaced signal, and neither side quite noticed the swap happening.

On the other side of that swap sits a recruiter who now has to sort a pile of applications that all look competent, because the tool generating them is the same tool everyone else is using. The honest response to that flood is to build a filter, so recruiting teams reach for their own AI screening tools, trained to catch exactly the patterns AI-written applications tend to produce. Which means the two tools are now effectively negotiating with each other, one generating text to pass a filter, the other refining itself to catch generated text, and the actual human beings on both ends of the transaction are increasingly bystanders to a contest between their own software. I have watched good recruiters describe this as exhausting in a very specific way, not the exhaustion of hard work, but the exhaustion of running fast on a treadmill that keeps adjusting its speed to match you.

The part that worries me most sits one layer further up, with the hiring manager who never sees any of this arms race directly. They see a dashboard. A shortlist, ranked, with scores that look precise because they carry decimal points. It is very easy, when you are busy and the number in front of you looks objective, to trust it more than it deserves. But a matching score measures resemblance to a pattern, not capacity to do the job, and the two only overlap by accident. What that score is quietly filtering out is judgment, specifically, because judgment does not show up as a keyword. It shows up in how someone describes a mistake, what they choose to leave out of an answer, the question they ask back when you give them an ambiguous brief. None of that survives being reduced to a match percentage, and a hiring manager who never sees the raw applications never finds out what got lost on the way to the shortlist.

I keep returning to a distinction I have leaned on in other contexts, because it applies here without much translation. A filter can tell you whether an application resembles the applications of people who succeeded before. It cannot tell you whether the person behind it will exercise judgment in a situation nobody has scored yet, which is the actual job in most senior roles worth having a real hiring process for. The industry keeps stacking more AI onto both sides of this exchange, hoping the arms race resolves itself if both filters get sophisticated enough. I don’t think it resolves. I think it just gets faster and more expensive to run, while the thing everyone was originally trying to measure keeps getting harder to see.

So the fix, if there is one, is not a better filter. It is choosing, deliberately, to reintroduce the things a filter cannot cheaply fake, at each point in the chain. A candidate who shows real work, a real referral, a direct message to an actual person, is doing something a hundred AI-tailored applications cannot replicate, because the whole value of it is that it did not scale. A recruiter who spends five real minutes in conversation instead of five seconds reading a score is buying back exactly the signal the volume destroyed. And a hiring leader who decides out loud which parts of the process stay human, and tells candidates plainly where AI is and isn’t involved, is doing the one thing that costs nothing except the willingness to say it, and pays for itself the first time it saves a good hire that the filter would have quietly dropped.

None of this is an argument against the tools. I use them, I recommend them, and I would not want to hire without them at the volume modern organisations operate at. But a tool that helps you sort volume is not the same as a tool that helps you see a person, and hiring is still, underneath every layer of automation we have added to it, the second job and not the first. The organisations that get this right in the next few years will not be the ones with the best screening model. They will be the ones who remembered, on purpose, to keep a human being somewhere in the loop whose actual job is to notice what the filter cannot.