AI Didn't Break Recruiting. It Audited It. Here's the Math.

Every hiring leader has a villain now. The overlay tool whispering into the IDE. The candidate who aced the system design round and then couldn't clone a repo on day one. The résumé that reads like it was assembled by a committee of one, and that committee was ChatGPT.

Here is the uncomfortable version. AI didn't break your hiring process. It ran the audit you were too scared to run yourself. It walked into a room that has been filthy since roughly the Eisenhower administration and flipped on the fluorescents. What you are calling a crime scene is just the lighting.

Let me show you the numbers, because the numbers have been screaming for twenty-five years while we kept the binoculars locked in the cupboard.

Start with predictive validity: the correlation between a hiring signal and how someone actually performs once they have the badge. Anything above 0.30 is doing real work. Schmidt and Hunter's 1998 meta-analysis, the most cited paper in the field, put structured interviews and general mental ability at the joint top of the hierarchy, both around 0.51. Unstructured interviews, the "let me just get a feel for the person" special, landed around 0.38.

Then in 2022, Sackett, Zhang, Berry, and Lievens re-ran the math. They found the field had been over-correcting for range restriction for half a century and quietly inflating everyone's favorite tools. Cognitive ability got revised down to roughly 0.31. And structured interviews? They came out on top of the entire hierarchy, the single strongest predictor of job performance at a mean operational validity around 0.42. Sackett called it the most important paper of his career. The reframe is brutal: the thing we treat as a formality is the most valid instrument we have.

Now the punchline that should keep you up at night. Most companies run the unstructured version. The one where a director "just wants to have a conversation." The one with no scoring rubric, no fixed questions, no defined competency, no decision made before the small talk starts. We took the lowest-validity method on the board and made it the load-bearing wall of the whole house.

That is not a candidate problem. That is a we problem.

Into that already-leaky funnel, pour the flood.

LinkedIn now processes something like 11,000 job applications per minute, a 45% jump in a single year, according to reporting by The New York Times. In the same window, the number of jobs actually posted fell about 10.6%. More applications. Fewer jobs. That gap has a name, and the name is generative AI. The average job opening now draws in the neighborhood of 242 applications, which pencils out to a rounding error of a chance for any one applicant.

The Willo 2026 Hiring Trends Report found 77% of hiring teams now regularly encounter AI-generated or AI-assisted applications. Only 37% of employers still consider credentials and learning history a reliable signal of talent, and 41% are actively walking away from résumé-first hiring. Read that twice. The document we built the entire top of the funnel around, the résumé, is being quietly declared unreliable by the people who read them for a living.

So what did we do? We fought bots with bots. LinkedIn's own research says 93% of recruiters plan to increase their use of AI in 2026. One side spray-and-prays with an auto-apply tool. The other side screens with a keyword scanner nobody validated. Two machines negotiating over a human being neither of them can actually evaluate. This is not hiring. It is a denial-of-service attack we launched on ourselves.

If you want the whole era in one kid, meet Roy Lee.

Twenty-one, Columbia, built a tool called Interview Coder, a hidden in-browser overlay that fed answers during live technical interviews without the interviewer seeing it. He used it, by his own account, to land interviews at the biggest names in tech. Columbia moved to suspend him. Instead of a career-ending scandal, he turned it into a founding story, rebranded the thing as Cluely, slapped on the tagline "cheat on everything," and raised $5.3 million in seed funding from real venture firms.

Here is the part nobody frames correctly. Cluely did not invent cheating. It productized a discovery: that a huge share of our "rigorous" technical screens were memorized trivia a machine could beat in real time. The overlay is not the disease. The overlay is the diagnostic. If a browser plugin can pass your interview, you were never testing what you thought you were testing.

And the kicker, the most on-brand ending imaginable. In 2026, Lee admitted he had publicly lied about Cluely's revenue numbers. The poster child for gaming the metrics got caught gaming his own metrics. That is not a footnote. That is the entire industry in a single human being: everybody optimizing the number, nobody minding the thing the number was supposed to represent.

Now price it out, because a broken process is not just embarrassing. It is expensive.

Time-to-hire runs somewhere between 42 days on the classic SHRM benchmark and the low 60s on Employ's 2026 data, depending on who you ask and how senior the role. Cost-per-hire sits around $5,475 for a non-executive role per SHRM's 2025 numbers, and jumps past $35,000 for executive seats, up more than 20% since 2022.

Then you get it wrong. The U.S. Department of Labor pegs the floor cost of a bad hire at roughly 30% of that person's first-year salary, and that is the conservative estimate. Once you fold in the manager babysitting the miss, the vacancy while you backfill, the team morale tax, and the lost output, credible modeling puts the real number anywhere from $30,000 to well past $150,000. Roughly three in four companies admit they have made at least one bad hire in the past year.

So let me summarize the value proposition. We are paying premium prices, waiting six to nine weeks, running the lowest-validity method available, and getting a coin flip with extra steps. You would fire a vendor who sold you this. Instead we defend it, because we have been doing it since before color television.

The fix is not a better cheating detector. It is not an AI arms race you will lose on cost. It is the least sexy word in the building: structure.

Same questions, every candidate. A defined competency for each question. A scoring rubric filled in during the interview, not a gut call reconstructed afterward. A work sample tied to the actual job, the thing they will really do, not a whiteboard puzzle from a 2011 algorithms course. And a decision rule you commit to before you ever shake a hand, so the "vibe" cannot quietly overrule the evidence.

Stop testing recall a machine will always win. Test judgment, taste, tradeoffs, the ability to be wrong out loud and recover. Those are the things no overlay can whisper, because they do not have a lookup answer. If your interview can be beaten by a browser plugin, the plugin did not expose the candidate. It exposed you.

The tool is not the threat. The threat is that we spent decades calling a vibe a process and a coin flip a meritocracy, and we billed full price for both.

AI did not change what good hiring requires. It just made the cost of pretending legible. It turned on the lights, and now we can all see the room. You can spend the next eighteen months yelling at the iceberg. Or you can pick up the mop.

The lights are not going back off. Build the thing that is actually worth passing.

Hi, I'm Brian Fink, Managing Partner of The ReWork Group. If you like how I think, be sure to check out my newest eBook + Audible, Talk Tech To Mini. Talk Tech To Mini skips the "what" of the technology and goes straight at the "how" of recruiting for it.