Zwischen

AI makes you average. Congratulations, if you weren't.

Generative AI is a spectacular floor-raiser and a lousy ceiling-raiser. Don't expect the labs to fix that. The floor is where the money is.

7 min read

You are probably world-class at one thing and a tourist in forty others.

The staff engineer who can smell a bad service boundary from the ticket title writes marketing copy like a tax form. The marketer who can read a room writes SQL like a ransom note. Ability isn’t a trait. It’s a position in a distribution, and you hold a different one in every domain you touch.

So “does AI make people better?” is the wrong question. Better from where?

AI raises the floor. It drags the bottom of every distribution toward the middle, does almost nothing for the top, and sometimes drags the top down.

Don’t wait for that to change. The floor is where the money is.

The floor moved

When a generative AI assistant rolled out to 5,172 customer-support agents, novice and low-skilled agents resolved 34% more issues per hour. Experienced, highly skilled agents got almost nothing (‘Generative AI at Work’, Quarterly Journal of Economics, May 2025).

The new hire gets the playbook the veteran spent years building. The veteran gets their own playbook back.

The ceiling didn’t

METR had 16 experienced open-source maintainers work real issues in codebases they’d contributed to for an average of five years. They expected AI to make them 24% faster. Afterward, they believed it had made them 20% faster. It had made them 19% slower (‘Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity’, METR, July 2025).

The tools have improved since. Forget the 19. Remember the gap: the people with the most context on earth about their own code couldn’t feel the tool costing them.

On open-ended, time-pressured business school exams, AI made low performers better and high performers worse until the two met in the middle (‘Leveling Up or Leveling Down?’, Academy of Management Learning and Education, 2026). The low performers copied the chatbot and skipped the analysis they couldn’t do. The high performers tried to digest everything it produced and lost the analysis they could do.

That’s the tax nobody prices in. Every answer the model hands you has to be read, judged and mostly thrown away. Below the median, that’s cheap, because its answer beats yours. At the top, you pay it to be told what you already knew.

THE AI REVIEWER WOULD LIKE A WORD

For the record, the evidence against us:

A median machine

Researchers compared 9,198 humans with 215,542 LLM responses on an established creativity task. On average, the two were close. At the extremes, it wasn’t close: humans varied far more and reached much higher in the right tail (‘A large-scale comparison of divergent creativity in humans and large language models’, Nature Human Behaviour, December 2025). Telling the model to play a genius helped up to a point, then fell apart.

The model is excellent at landing in the fat middle of a good distribution and bad at the rare extreme answer that makes someone exceptional. Exceptional is the only thing your best people are paid for.

Our bet on why: models learn from what people wrote down, and what people wrote down is overwhelmingly ordinary – a thousand tutorials for every brilliant design. Then post-training asks people which answer they prefer, and a vote rewards what most people recognize. It’s the same argument we made about coding agents in The Homer.

Follow the money

Here’s the pitch. Dario Amodei’s “Machines of Loving Grace” imagines AI that could “compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years,” with the “elimination of most cancer” and a “doubling of the human lifespan” along the way (‘Machines of Loving Grace’, Dario Amodei, October 2024).

Here’s the product. OpenAI’s own economists found ChatGPT had reached about 10% of the world’s adults, and more than 70% of messages weren’t about work. Nearly 80% of conversations fell into three buckets: practical guidance, seeking information and writing (‘How People Use ChatGPT’, NBER, September 2025).

Everyone is a tourist in almost every domain, so “below the median at the thing I’m asking about” isn’t a segment. It’s nearly everybody, nearly all the time. How to word the email to the landlord. What the rash is. Why the spreadsheet formula broke. Move any of them from the 20th percentile to the 50th and they feel it that afternoon. That feeling renews subscriptions.

At the top, the math flips. The right tail of any field is a sliver of people: the hardest users to impress, the most expensive to build for, and, if METR is any guide, not reliable judges of whether the tool is helping. A customer who pays and believes it’s working is a customer. Nobody has to build the moon-shot tool to keep them.

The moon shot is how you raise the money. The floor is how you make it.

You can’t grade the tail

Every improvement needs a grader, and graders get expensive fast as you move right. An experienced professional can tell you in minutes whether a memo is competent. Only exceptional people reliably recognize exceptional work, there are almost none of them, and some of it can’t be graded for years. A cancer cure is only a cure once the trial reads out.

When OpenAI built a benchmark for professional-grade work, GDPval drew its tasks from “industry professionals with an average of 14 years of experience” (‘GDPval’, arXiv, October 2025). That’s a serious bar, and the wrong one for this question. Fourteen years tells you someone is experienced, not that they’re in the top 1% of anything. Everyone has worked with the 14-year veteran who is exactly as good as they were in year four.

We don’t know who grades any given lab’s training feedback. We do know the cost curve. Nobody has to decide to aim at the middle for the incentive to point there. So when a benchmark says a model matches experts, ask which ones.

THE GRADER
Experience is a credential. It isn’t evidence of the right tail.

Where expertise wins

The counter-case is real. In a study where AI closed about three-quarters of an education gap on a workplace-style business problem, the chat logs still split: lower-education participants leaned on the AI for substantial help, while higher-education participants used it somewhat more effectively (‘Does Generative AI Narrow Education-Based Productivity Gaps?’, NBER, February 2026). One group was borrowing capability. The other was steering it.

That’s the second force. Substitution wins when the task has a known answer near the middle and you don’t have it. Leverage wins when output is cheap to judge and judgment is your edge. AI goes negative at the top when the work is open-ended, the output is voluminous, and reading it crowds out the thinking.

THE TEST
The question isn’t how good you are. It’s whether you can tell when it’s wrong.

The median, delivered

None of this makes AI useless to strong people. It makes it a product that wasn’t built for them. Where you’re a tourist, it’s instant competence. Where you’re exceptional, it’s a fast, confident colleague who is reliably average.

The labs will keep promising the cure for cancer, and they’ll keep shipping the landlord email. Both are rational. Only one is the business.

Use it where you’re below the median. Keep it on a short leash where you’re not.

Further reading.

The answer is 14. Now what?

AI will build you 87 flawless charts before your coffee is cold. The old data hands knew the only question that mattered came after the number, and nobody is asking it anymore.

8 min read

All posts