The best people I work with are quietly getting worse at the part of their job that matters most.

Not the ones who refuse to touch AI. The ones who use it constantly.

The moment it stopped being a hunch was a Tuesday call with a head of strategy at a consulting firm. He was walking me through a market-entry recommendation his team had built, and it was clean. Sized market, three scenarios, a clear pick.

I asked him why they ruled out the second scenario. He paused, scrolled, and read me back the model's reasoning almost word for word. So I asked a different way. What did he think. There was a long silence, and then he said he would have to go back and look.

This was a man who, two years earlier, would have argued me through that decision tree from memory and enjoyed it. He had not gotten lazy. He was working more than ever. He had just quietly stopped doing the one part of the job that used to be his edge, because the model did it faster and the answer looked right. The muscle had gone slack and he had not noticed, because nothing on the outside looked any different.

Here is what I have come to believe. There are two ways to use AI, and from the outside they look identical.

One uses the model to think harder. You argue with it. You make it show its work. You catch it when it is confidently wrong, because you still know enough to catch it.

The other uses the model to think less. You ask, it answers, you ship. It feels like speed. It is actually a handover.

The gap between those two is tiny in the moment and enormous over a year.

Ethan Mollick wrote about this last week, and he put numbers on something I had only been able to describe as a feeling.

He points to a study of about a thousand high school students in Turkey learning math. One group got plain ChatGPT. One group got nothing. The AI group did their homework better and felt like they were learning more. Then they sat a test with no AI, and scored worse than the students who had struggled through without it. The model had been handing them answers. Answers do not build the muscle. The struggle does.

Then look at the flip side. A second study, many of the same researchers, a five-month coding course across ten schools in Taipei. This time the AI was set up to tutor: to hand students a tailored sequence of problems instead of solutions. Those students scored meaningfully higher on an exam they took with no AI in the room at all. Same technology. Opposite result. The only thing that changed was whether the AI did the work or made the student do it.

That is the whole thing, and it is not a school problem.

The part executives should sit with is Mollick's own research. He and his colleagues ran an experiment on 758 consultants at Boston Consulting Group. Half got access to GPT-4. On most tasks the AI users crushed the others: faster, better, more polished. Then the researchers slipped in a problem the model was known to get wrong. On that one, the consultants with AI were significantly more likely to get it wrong too. The model gave them a confident, authoritative answer. These were elite people. Most of them took it.

His Wharton colleagues have a name for what happened: cognitive surrender. People stop wrestling with the problem and let the machine carry it, even when the machine is wrong.

I keep coming back to that phrase because it describes something I see every week and could never name.

Here is why this should worry you more than any headline about AI replacing jobs.

The risk is not that AI does your team's work. The risk is that your team slowly forgets how to do it, and you do not find out until the day the model is confidently, expensively wrong and nobody in the room is sharp enough to catch it.

And the newer the tool, the faster this happens. The old chatbots made you work. They lost the thread, they made obvious mistakes, they forced you to stay in the loop. The agentic tools are built to remove friction. They just do the thing. But that friction was doing quiet work: it was keeping your people in contact with the problem. Strip it out and you get a smoother workflow and a duller team, and the second part does not show up on any dashboard.

The move is to change what your team hands you, not what they hand the model. Stop accepting the polished recommendation as the deliverable. Ask for the recommendation and the one assumption that, if it is wrong, breaks the whole thing, in their own words, with the model closed. You are not banning the tool. You are making the thinking the thing you inspect, instead of the output. Do that consistently and people start keeping the hard part for themselves again, because they know they will have to defend it.

What I tell the leaders I work with is this. Stop measuring whether AI made your team faster. Everyone is measuring that. Start watching whether your people can still defend the answer when you take the AI away.

Run the test now and then. Take a recommendation that came out of a model and ask the person who delivered it to walk you through the reasoning with the screen closed. You will learn very fast who is using AI to think and who has handed it the thinking. It is not a gotcha. It is a fire drill for a skill you cannot afford to let atrophy.

Mollick is careful, and I want to be careful too. This is not an argument against using AI for writing, or analysis, or any of the work it is genuinely good at. I let AI pressure-test almost everything I make. I am happy to never do mental arithmetic again. The point is not abstinence. The point is that we are setting the defaults right now, the AI companies designing for the frictionless answer, employers deciding what counts as "using AI well," everyone teaching some version of AI literacy, and we are doing it mostly without thinking. Once a generation of people builds its habits around handing over the hard part, those habits will be very hard to reverse.

So the people pulling ahead are not the ones using AI the most. They are the ones who have decided, on purpose, what they refuse to hand over.

That decision is the actual skill now. Not prompting. Not tooling. Knowing which parts of the work are load-bearing for your own judgment, and protecting those even when the model is reaching out to take them off your plate.

The frictionless answer is always there. Choosing not to take it is the job.