Everyone Says Hire for AI Skills. That's Not the Scarce Skill Anymore.

Every hiring article published in 2026 tells you to screen for AI skills. Ken Ono, a mathematician with more than 30 years of research behind him, sat down determined to find one math question ChatGPT couldn't answer. He couldn't find one (in an interview on the EO Việt Nam channel, September 2026). His conclusion wasn't that AI had made mathematicians obsolete. It was that he'd been asking the wrong question, and most companies evaluating talent right now are making the same mistake.

This post is for founders and HR leaders making hiring, review, or promotion decisions in the next two quarters, at a company where AI already touches day-to-day knowledge work. It's not a claim that AI skills don't matter at all, or that credentials are worthless. Most performance systems still reward raw knowledge and output volume. That's exactly the axis AI is collapsing, and the replacement axis is harder to see.

What is every hiring guide telling you to look for right now?

AI fluency is a hiring signal that has gone mainstream fast. One widely quoted line from a tech CEO captures the mood (Entrepreneur, 2026): "hiring someone today without assessing their AI skills is like hiring someone in the 1990s without checking if they could use the internet." That's a fair bar. It's also already becoming table stakes rather than a differentiator.

Here's the problem with stopping there. Not a small one.

• If AI fluency is the finish line, you're optimizing for a skill that gets cheaper every quarter as the tools get easier to use.
• A candidate who lists five AI frameworks on a resume has told you about tool exposure, not about their ability to make a good call under uncertainty (Future of Talent, 2026).
• Managers already report struggling to tell who on their team is actually good versus who just produces a lot of AI-assisted output that looks good (Chattanooga Times Free Press, June 2026).

None of that means AI skills training is wasted. It means the finish line moved, and most companies haven't noticed.

What did a mathematician's failed attempt to stump ChatGPT reveal about talent?

Ono's real project isn't proving AI is smart. It's finding the next Ramanujan, referring to Srinivasa Ramanujan, the self-taught Indian mathematician whose work was overlooked by formal institutions before a British mathematician recognized it. Ono has spent his career looking for people with that same shape of ability: original mathematical judgment that doesn't show up on a standard test.

His point, translated out of academia: once a machine can hold and produce more raw knowledge than any human alive, "what do you know" stops being the differentiator. "Do you know who you are," meaning your own judgment, taste, and the specific questions only you think to ask, becomes the scarce thing.

That's not a soft, feel-good idea. It's structural. And it cuts against how most companies already run reviews. Standard evaluation systems, whether school grades or standard-issue performance reviews, are built to measure output against a known rubric. A rubric is exactly what AI is now best at satisfying. The person whose edge is an original angle nobody assigned them will often score average on a rubric built before AI existed. That changes the moment their judgment turns out to be the reason a project didn't fail.

Why do your current hiring and review signals miss this kind of person?

Most companies still hire and promote against three pre-AI signals, and each one is losing predictive value at a different pace.

1. Credentials. A degree or certification tells you what someone was taught, not what they do when the answer isn't in the training material.
2. Tenure. Years in a role used to proxy for accumulated judgment. It still can, but AI now lets a two-year employee produce the volume that used to take five.
3. Output volume. This is the fastest-collapsing signal. If a task is measurable purely by units produced, AI can already do more of it than your best employee, which means rewarding volume rewards whoever adopts the tool fastest, not whoever has the best judgment about when to trust it.

None of these are useless. But when all three point the same direction, companies that promote whoever produces the most AI-assisted output start actively selecting against the Ramanujan-shaped employee. That's the one who ships less volume because they spent an extra hour asking whether the AI's answer was actually right for your specific business, not just plausible-sounding.

What does a practical "judgment check" look like inside a company?

Here's a concrete framework, adapted from Ono's search for overlooked mathematical talent into something a manager can actually run this quarter.

Step 1: Diagnose. Pick one recent piece of AI-assisted work from your team, a report, a proposal, a piece of code. Don't ask who produced the most of it. Ask who caught the thing in it that was wrong.

Step 2: Redesign the review question. Replace "how much did you ship" with two questions in your next 1:1 or performance review:

• What's one thing the AI got confidently wrong that you caught, and how did you know?
• What's a question about our business you've been sitting on that nobody assigned you?

Step 3: Watch for the tell. The employee worth nurturing here isn't always the fastest talker in the room. Often it's the one who pauses, admits uncertainty on the parts that are actually uncertain, and has a specific, narrow answer instead of a broad confident one.

In my own hiring interviews, I like to put candidates inside a scenario close to a real problem and watch how they reason through it, not just what they conclude. The hires who turned out best were rarely the fastest talkers. They were the ones who noticed a small detail everyone else skipped past, then pushed back on the premise. They'd ask why we were solving it this way, or what they were missing before committing to an answer. That instinct, catching the overlooked detail and questioning the frame before answering, is exactly where AI still falls short. It's fast at mimicking the shape of human reasoning, but it misses the small detail that, more often than not, turns out to be the one that makes the real difference.

Step 4: Protect their time. If you find this person, the worst thing you can do is measure them on the same output metric as everyone else. Give them the ambiguous problems, the ones without a rubric yet, and let their score on the boring metrics stay average.

A quick gut check before you run this: is this just a rebrand of "hire smart people"? Not entirely. Close, but not entirely. Smart-people proxies (test scores, brand-name schools) correlate with the ability to learn fast within an existing rubric. What Ono describes is closer to taste, the ability to notice which question is worth asking before anyone hands you the rubric. Those two things overlap, but they're not the same, and companies that only screen for the first will keep missing the second.

Frequently asked questions

Should I stop hiring for AI skills entirely?

No. AI fluency is still worth screening for, but treat it as a baseline requirement, not the differentiator. In 2026, most candidates who apply for knowledge-work roles already use AI tools daily. The real signal is what someone does with the AI's answer, not whether they can produce one.

How do I test for judgment in an interview if there's no test for it?

Give the candidate an AI-generated answer to a real problem from your business and ask them to find what's wrong with it, then defend their fix. Their reasoning process, not their final answer, is what you're evaluating.

Isn't this just a fancy way of saying "hire smart people"?

Not quite. Judgment in this sense is closer to taste plus self-direction, the ability to notice a problem worth solving before anyone assigns it. That doesn't always correlate with the credentials or test scores companies still use as smart-people proxies.

What's the business risk of getting this wrong?

You keep promoting people whose only edge was pre-AI output speed, which AI now erases. The person with the best judgment on your team gets passed over because it never showed up in a metric, and eventually leaves for a competitor who noticed.

Key takeaways

• AI fluency is becoming table stakes in hiring, not the differentiator it was in 2024–2025.
• Credentials, tenure, and output volume are all losing predictive value at once, and they're the three signals most review processes still lean on.
• The scarce trait now is judgment: knowing when an AI output is wrong and having the self-directed curiosity to ask questions nobody assigned.
• You can check for it this quarter with two review questions and one exercise, no new HR software required.
• Protect the person who has it by keeping them on ambiguous problems, not by holding them to the same output metrics as everyone else.

I'd genuinely like to know if this framework holds up outside a research-lab or founder context, especially in larger orgs with formal HR systems built entirely around the old signals. If you've tried something similar and it broke, tell me where.

---

Credit: Photo by Tima Miroshnichenko on Pexels