Most of us know the feeling. You ask the chatbot a question, the answer comes back fluent and confident, and you ship it. You meant to check the numbers, push back on the framing, run the obvious counterargument. You didn't, because the answer felt right and the day was already long. Researchers at the University of Pennsylvania have a name for this now: cognitive surrender.

In their experiments, AI users accepted the system's reasoning 93% of the time when it was accurate — and went down with the ship when it wasn't. Their performance tracked the model's quality. They had stopped thinking on their own behalf and started travelling with the machine instead.

The reason this matters for anyone running or working in a knowledge business is straightforward enough. Cognitive surrender isn't a failure of access to AI. It's a failure of judgement, and judgement is the thing that's actually being paid for. The professionals and the firms who use AI without giving up the harder bit of thinking are the ones who'll outperform over the next five years; the ones who don't are buying speed today at the cost of their own future value.

The data is starting to land

Three pieces of evidence sit on top of each other now, and they tell a consistent story.

Start with the Pennsylvania research, recently picked up by Time and Ars Technica. The bit that should worry employers isn't the headline finding so much as who it applies to: on measures of fluid intelligence, the people most inclined to surrender to AI were not the brightest in the room. Brighter people overruled faulty AI responses more often. Everyone else just followed the machine.

A separate 2025 study by Microsoft and Carnegie Mellon makes the mechanism harder to look away from. Surveying 319 knowledge workers using generative AI at work, the researchers found that high confidence in AI is associated with less critical thinking, not more. Routine work is the gym where we practise judgement; when you mechanise the routine and only call the human in for exceptions, you arrive at those exceptions with cognitive muscles that have atrophied and become unprepared. The phrase isn't ours; it's theirs.

Anthropic's 81,000-person qualitative study, run across 159 countries earlier this year, sits on top of the other two. 16.3% of respondents — a meaningful slice of Claude's own users — volunteered cognitive atrophy as a worry without being prompted. Of the people who said they were worried about it, 46% reported they had already seen it happen to themselves. They weren't anxious about a future risk; they were confessing to a present one.

How knowledge workers actually use AI

Whether AI makes us smarter or dumber depends entirely on the relationship we've built with it. I'm not sure anyone really knows yet how that shakes out across a whole career — but at the level of individual habits, we can see two patterns clearly enough now.

Most professionals are using AI in one of two ways, and they produce wildly different outcomes.

Mode one is AI as oracle. You give it a question, it gives you an answer, you ship the answer. The relationship runs on trust: the AI is treated as a colleague who already knows what you don't. Your client, your boss, your team gets the AI's work with your name on it. You get the speed dividend. The hidden cost is that you stop developing the thinking that used to be your job. Every task you outsource is a rep you don't do. Six months in, you're faster but you're also weaker.

The healthier posture treats AI as a sparring partner. The shape of the interaction is the same: you ask, it answers, but afterwards you push on what came back. You ask whether it considered the obvious counterargument. You test where it might be wrong. You bring in the context the model doesn't have. The friction is what does the useful work, and you end up both quicker than before and sharper at the underlying judgement, because friction itself is a form of practice. It's the same posture I argued for in the 75/25 split when looking at AI coding tools: they're brilliant for most of the work, but the 25% where they're wrong is exactly where your judgement has to stay live. Anthropic found that 91% of people who reported learning benefits from AI had collaborated with it in this second mode. People in the first mode reported atrophy.

What separates the two postures isn't access or training; it's a willingness to keep doing the harder bit of the job once a machine can do the easier bit for you.

What this means for your business

If you employ knowledge workers, you've got a problem most leaders aren't tracking. Your team is getting through more work, and the dashboards look healthy. But underneath that, the people who do the work are slowly losing the cognitive habits that made them valuable. The reckoning shows up later, when a client asks a question the AI got wrong, when a decision needs to be made on incomplete information, when the model is down and the deadline isn't. This is also why AI literacy for managers matters more than vendor demos ever will: you can't manage what your team has stopped doing if you can't tell good AI output from confident-sounding nonsense.

The same erosion plays out inside software teams as cognitive debt, where the group ships code faster than anyone can still explain it.

This isn't an argument against using AI. Used well, AI delivers genuine productivity gains, and refusing to engage with it leaves you behind anyone who does. The harder thing to admit is that most of the productivity gain is being captured in a way that erodes the asset producing the gain. We're spending the cognitive principal as well as the interest. How recoverable any of this proves to be, nobody can yet say. Some habits may bed in fast and prove hard to reverse; others might shake off within weeks of disciplined use.

The professionals who'll win the next decade are the ones who use AI without surrendering to it. They're rarer than the hype would suggest.

How to keep your thinking sharp

Staying in the sparring-partner camp — with your judgement still intact in five years — comes down to a small number of habits. None of them are clever. All of them require effort, which is exactly why most people skip them.

  1. Form your own view first. Before you ask the AI to draft something, write down what you'd say if the AI didn't exist. A line, a paragraph, a rough position. Then let the AI help. The first version doesn't have to be good; it's got to be yours. Skipping that step is what hollows out the muscle over time.
  2. Treat the AI's output as a draft to argue with. Read it the way a sceptical editor would. Find one thing it got wrong, one thing it overstated, one thing it missed. If you can't find anything, you weren't reading carefully — every AI output of any length contains at least one of those.
  3. Refuse to ship anything you don't understand. Obvious in theory, broken constantly in practice. If the AI gave you a number, a name, a framework, or a recommendation that you couldn't reconstruct without it, you don't yet know enough to use it. Either learn it or remove it.
  4. Keep doing some of the work by hand. Not all of it — that defeats the point. But enough that the muscles stay warm. Write the first email yourself. Sketch the analysis on paper. Make the call without prompting the chatbot first. The reps still matter.
  5. Audit your own atrophy. Once a quarter, do a task you used to do well, without AI. Notice where you've slipped. That gap is your honest measure of how much you've outsourced, and it tells you what to take back.

The bottom line

AI was sold to knowledge workers as a way of doing more of the work they were already doing. For a significant share of users — the ones who use it most heavily and trust it most readily — it's becoming something different: a way of doing less of the thinking that the work was actually about.

The competitive advantage no longer goes to the people with the best AI tools. Everyone has those. It goes to the people who can use those tools without surrendering the judgement that made their work worth paying for. That's a smaller club than the productivity literature wants to admit, and it's the one worth being in.