The Uncomfortable Conversation Every CHRO Needs to Have About AI Hiring

The conversation you keep postponing is the one where you admit, out loud and to your CEO, that most of your screening spend is now measuring the candidate's tool and not the candidate. Your ATS reads text a model wrote. Your take-home grades a model's output. Your "must have AI skills" line sorts nobody, because everybody clears it. You already suspect this. So does your head of talent. Nobody has said it in the room yet, because saying it means admitting the last two years of process investment bought you a more expensive way to measure the wrong thing.

I am writing this to you as a peer, not a vendor. You do not need another tool. You need to have the meeting.

The short version:

• Your hiring apparatus was built to measure the person through what they produced. AI severed that link. The ATS, the take-home, the competency screen now read the quality of the candidate's model, not the candidate.
• "Must have AI skills" is the tell. It feels rigorous and current, and it sorts nobody, because in 2026 fluency with the tool is universal. A screen everyone passes is a formality wearing the costume of rigour.
• The spend is theatre, and the function half-knows it. The people running your assessments can feel the artifacts stopped meaning anything. What they lack is permission to say so, which is a thing only you can grant.
• Fixing this costs you nothing in software and everything in comfort. You change what you measure, not what you buy. The one signal left is what a candidate does when the model is confidently wrong in front of them.

---

The apparatus was a proxy machine, and the proxy broke

Walk back to first principles, because that is where the honest version of this conversation starts. Your entire function is a proxy machine. You cannot observe capability directly, so you built instruments that stood in for it. The CV stood in for a track record. The take-home stood in for how someone works. The competency screen stood in for judgment under a real problem. Each worked for one reason: producing the proxy used to cost about what having the real capability cost. To submit a sharp strategy memo, you had to think strategically. The artifact was expensive to fake, so its presence was evidence.

AI did not make your candidates more capable. It made the proxies almost free to produce while leaving the underlying capability precisely as rare as it always was. That is the whole event, stated plainly. Cheap artifact, rare capability, and a gap between them your process falls straight through. The instruments still return numbers. The numbers no longer mean what the instrument was built to make them mean.

This is why the uncomfortable feeling in your talent team is not incompetence. It is accuracy. They are watching a dial that used to track capability and now tracks the candidate's willingness to open a browser tab. They know the dial is lying. They keep reading it because the process says to, and because nobody senior has said the dial is broken.

"Must have AI skills" is the sentence that gives you away

Look at your last ten job descriptions. I would put money on the phrase "must have AI skills" or a close cousin appearing in most of them. Every hiring manager has written the line. Almost none of them can tell you what it screens for, because it screens for nothing that separates one candidate from another.

AI-related skills now appear on roughly 2.5% of US job postings and climbing, and around 4.2% of entry-level roles, per PwC's AI Jobs Barometer. The line is spreading through your job boards faster than anyone can define it. And "AI skills" names one thing only: fluency, the ability to make the tool produce output. In 2026, fluency is what a teenager has by default and a course teaches in an afternoon. A requirement met by everyone who applies is not a filter. It is decoration on a form.

The deeper problem is the one I laid out in "must have AI skills" means nothing: the phrase points at something real and grabs the wrong thing. You want direction, the capacity to steer the model, override it, and catch it when it is confidently wrong. What you wrote down was fluency, because fluency is the part you can see and name. So the requirement sorts on the axis where everyone scores full marks and stays silent on the axis that predicts the hire.

You are measuring the tool. Here is the proof.

Say a candidate leans on AI hard and checks it almost never. Picture their submission. More finished than everyone else's. That is the part that should keep you up.

Wharton researchers Steven Shaw and Gideon Nave, in a study titled Thinking, Fast, Slow, and Artificial, ran three experiments with 1,372 people on reasoning tasks. Accuracy climbed from 46% with no AI to 71% when the AI was right. Then, when the AI was wrong, accuracy fell to 31.5%, below the 46% people managed with no help at all, and participants followed the wrong answer around 80% of the time while growing more confident as they got it wrong. Shaw and Nave call it cognitive surrender.

Read that as a hiring instrument and it inverts everything. The most polished artifact in your pile is the one produced with the least thought, by the candidate who surrendered hardest. Your take-home does not reward judgment. It rewards a confident model and a candidate willing to trust it. You are not grading people against each other. You are grading their models, and awarding the top score to whoever argued with theirs the least.

That is the sentence to bring to your CEO. Not "AI is changing hiring." Every deck says that. The sentence is: our best-scoring candidates and our least careful candidates are now the same candidates, and our process cannot tell them apart.

The comfortable belief	The uncomfortable reality	What it forces you to change
Our ATS screens candidates	It screens text a model wrote to satisfy another model	Demote the CV to routing, not verdict
"Must have AI skills" raises the bar	Everyone clears it; it sorts nobody	Requirement rewritten around direction, not fluency
The take-home shows how they work	It shows how their model works	Move the assessment into a live, unpredictable moment
A polished submission signals a strong candidate	Polish now correlates with surrender to the tool	Score the wrong-answer catch, not the finish
More assessment tooling means more rigour	More tooling measures the tool more precisely	Spend on what you measure, not what you buy

The spend is theatre, and everyone in the function knows it

Here is where the conversation gets genuinely uncomfortable, because it stops being about candidates and starts being about your own budget.

You have spent real money on the assessment stack. The ATS licence, the structured-interview platform, the take-home grading rubric, the psychometric add-on someone sold you in 2024. All of it assumed the artifacts carry information about the person. That assumption is dead, which means a large share of that spend is now theatre: an expensive performance of rigour that produces a number nobody trusts and everybody keeps quoting.

Your team feels this. The recruiter who reads a flawless take-home and gets a flat hire three months later feels it. The hiring manager who cannot say why the "strong on paper" candidate cannot hold a live discussion feels it. Nobody escalates, because escalating means questioning a system you signed off on, and that carries political cost. So the theatre continues, funded and quietly distrusted by the people running it. What a company looks like after a year of this is not abstract; I traced it in 12 months after hiring for usage. The velocity charts stay green while the ability to catch a wrong answer erodes underneath.

It persists for a structural reason, and it is not your team's fault. You cannot manage what you cannot measure, and you cannot fairly ask a recruiter to measure direction when nobody has defined it, scored it, or trained them on it. That gap is the real subject of you can't train what you can't measure. The theatre fills the space where a real measurement should be.

The fix is a demotion, not a purchase

The instinct, when a CHRO accepts all this, is to ask which product fixes it. That instinct is the trap. The problem was never a missing tool. It was measuring the wrong layer, and no tool measures the right layer for you. The fix has two moves, and neither has a licence fee.

First, demote every pre-collected artifact to routing information. The CV tells you where to look, not who to hire. The portfolio starts a conversation, it does not end one. The take-home, if you keep it at all, becomes a prompt for a live discussion, never a score on its own. You stop treating anything produced offstage as evidence, because offstage is exactly where the tool does the work.

Second, move the real assessment to the one place AI cannot follow the candidate: a live moment, on a problem they could not pre-cook, where the model is confidently wrong in front of you. Hand them a real task with an AI output already attached, clean and plausible and quietly incorrect, and say nothing about the error. One person accepts the gift and builds on it. Another gets an itch, checks the load-bearing claim, finds the crack, and rebuilds the part that was rotten. That behaviour cannot be faked, because faking it would require the exact judgment you are testing for. This is the full case I made in stop hiring AI users, start hiring AI Operators: fluency is table stakes, direction is the hire, and only one shows up under a wrong answer.

Notice what changed and what did not. The spend can go down. The rigour goes up. You are measuring the person again instead of their instrument, and you are doing it with less software, not more.

Have the meeting

The conversation you have been avoiding has three sentences in it. Our screening measures the tool, not the person. "AI skills" as a requirement sorts nobody. Fixing it means changing what we measure, not buying another platform. Say those three sentences to your CEO before a bad hire says them for you.

None of this is a failure of your function. Your instruments were sound for a hundred years and one technology retired them in eighteen months. The failure would be knowing the dial is broken and continuing to read it because reading it is what the process says to do. You are the only person senior enough to stop the theatre and fund the measurement that replaces it.

Ivanooo built the AI Operator Profile to make that measurement repeatable: to put a candidate in front of a confidently wrong machine and read the one signal AI cannot produce on their behalf. If "must have AI skills" is on your job descriptions, this is the thing you were reaching for and could not name.

---

Frequently asked questions

Why is my hiring process measuring the tool instead of the candidate? Because every artifact it collects was a proxy that only worked when producing it cost about what having the real capability cost. AI made the artifacts nearly free to produce while leaving the capability rare. So the CV, the take-home, and the writing sample now report the quality of the candidate's model rather than the candidate.

Isn't "must have AI skills" a reasonable requirement in 2026? It reads as reasonable and screens for nothing that separates candidates. "AI skills" names fluency, and in 2026 nearly everyone who applies has it. AI-related requirements now appear on roughly 2.5% of US postings per PwC's Barometer, spreading faster than anyone can define them. A requirement everyone meets is not a filter. What you actually want is direction, which the phrase never measures.

How do I explain this to my CEO without sounding like I'm attacking our own process? Frame it as instruments, not incompetence. Your assessment stack was sound for decades; one technology retired it in eighteen months. The sharp version is one sentence: our best-scoring and least-careful candidates are now the same people, and our process cannot tell them apart. That is a factual observation about measurement, not a confession of failure, and it invites a fix rather than blame.

Does fixing this mean buying new assessment software? No, and reaching for a product is the trap. The problem is that you are measuring the wrong layer, and no tool measures the right layer for you. The fix is a demotion and a relocation: demote every pre-collected artifact to routing information, and move the real assessment into a live, unpredictable moment where the model is confidently wrong in front of the candidate. That costs comfort, not licence fees.

Why does my team keep running assessments they seem to distrust? Because distrust without permission produces theatre. Your recruiters can feel that a flawless take-home no longer predicts a strong hire, but escalating means questioning a system leadership signed off on, which carries political cost. You cannot fairly ask them to measure direction when nobody has defined or scored it either. Granting permission to change what you measure is a thing only the CHRO can do.

What is the one signal that still separates candidates? What a candidate does the moment the machine is confidently wrong. Hand them a real task with a plausible-but-incorrect AI answer attached and watch whether they catch it. Wharton's Shaw and Nave found people followed wrong AI answers around 80% of the time while growing more confident, so polish now signals surrender, not judgment. The wrong-answer catch cannot be faked, because passing it requires the exact capability you are hiring for.