Thinking Next to a Machine

Where is the wisdom we have lost in knowledge?
Where is the knowledge we have lost in information?

T.S. Eliot, Choruses from "The Rock," 1934

Eliot is describing a slide in the wrong direction, from lived understanding down to a pile of fragments. That is exactly the slide a fluent machine can paper over, because it sounds like it knows things while it is really working on correlation.

Judgment is a practiced skill

Aristotle had a name for the thing I keep noticing is missing. He called it phronesis, usually translated as practical wisdom. It is not being correct in the abstract. It is the hard-won ability to work out what to do here, in this specific situation, with these particular competing goods pulling against each other. He was blunt that the young cannot really have it yet, because it grows out of experience, and experience takes years to accumulate.

He also drew a line between practical wisdom and plain cleverness. Cleverness is good at hitting whatever target you are handed. Practical wisdom is about choosing the right target to begin with. A model can be enormously clever. It has read more accounts of wise decisions than any of us ever will. It has just never had to make one and then live inside the consequences.

What the machine cannot reach

The philosopher Hubert Dreyfus spent a career arguing that human intelligence is not a rulebook running on a pile of facts. It is embodied. We lock onto what matters in a situation, and quietly ignore a thousand things that do not, because we are already standing inside a world that means something to us. The machine is great at generating plausible next sentences. Dreyfus points at the question underneath that. How does a situation come to register as mattering at all?

Michael Polanyi put the same idea into five words in 1966 that have outlasted most of what surrounded them.

We can know more than we can tell.

Michael Polanyi, The Tacit Dimension, 1966

Language models are trained on the part we managed to write down. The tacit part, the felt sense of what is relevant that an experienced person carries around and could never fully explain, was never in the text in the first place. What the model picked up is the shadow that knowledge throws onto language, not the knowledge itself.

What the evidence actually shows

Here I want to slow down, because it is easy to grab a frightening headline and run. The research does not say the tools make us stupid. It says something more careful and, to me, more interesting. Tools move effort around. They can free up attention, and they can also quietly retire the practice that kept a skill in shape.

The first hint came from memory. In a 2011 study in Science, Betsy Sparrow and her colleagues found that when people expect to be able to look something up later, they hold onto the fact itself less and hold onto where to find it more. It was a short lab study with 46 students, so it shows effort moving around, not a long slow decline. But anyone who has stopped memorizing phone numbers knows the feeling.

Navigation is the cleaner example. In a 2020 study, Louisa Dahmani and Veronique Bohbot found that heavier lifetime GPS use went hand in hand with worse spatial memory when people had to find their own way, and a small follow-up hinted the gap widened over three years. A separate brain-imaging study found the navigation parts of the brain barely stir when you are just following the turn-by-turn voice. The automation does not take the trip away from you. It takes away the active work that kept the underlying skill alive.

The most on-point study is recent. In a peer-reviewed 2025 paper, researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 real tasks. The line worth sitting with is this. The more people trusted the AI, the less critical thinking they reported doing. The more they trusted themselves, the more they did. The thinking did not disappear. It shifted, away from producing the first draft and toward checking, combining, and supervising whatever the machine produced. It is self-reported, so read it as a picture of habits rather than a brain scan.

There is also a much-discussed MIT preprint that wired people up with EEG headsets while they wrote essays, and found the chatbot group showed the weakest brain connectivity and had trouble quoting work they had finished minutes earlier. It is a striking result. It is also a preprint, not yet peer reviewed, with a small sample and a published critique already attached, so I am holding it loosely. The honest read across all of it is the same. Offloading moves effort around. Whether it wears the skill down depends on what you do with the time it frees.

We have been here before

None of this worry is new, which I find oddly comforting. The oldest version is in Plato. Socrates fretted that writing would weaken memory and leave people with the appearance of wisdom instead of the real thing, an understanding propped up on outside marks rather than held inside. He was half right. Writing did push memory outside the skull. It also made possible nearly everything we now call thinking, including the very dialogue that carried his complaint down to us.

The printing press got the same reception. My favorite example is a Venetian editor named Squarciafico, who in 1477 grumbled that the new flood of books would make men less studious. Swap "books" for "answers" and it could be a post today. Thoreau, eyeing the telegraph in 1854, said the thing that has aged best of all. We were in a great hurry to wire Maine to Texas, he noted, even though Maine and Texas might have nothing important to say to each other. New tools, he warned, are often improved means to an unimproved end.

The pattern is always the same, and it is not the lazy version where people panic and are simply wrong. It is subtler than that. The critics usually got the losses right and the timeline wrong. Something real does get rearranged, in memory, in attention, in habit. And new tools tend to grow new ways of coping right alongside the losses. Both happen at once, and which one matters more is only clear a generation later.

Where we actually are now

Here is where I stop reasoning and just look at numbers. The broadest recent measurement comes from a 2025 study by KPMG and the University of Melbourne, which surveyed 48,340 people across 47 countries. The headline is a gap. People reach for these tools far more readily than they trust them.

Figure 1

Use runs well ahead of trust

Two-thirds of people use AI regularly, and most are fine with it being used. Fewer than half are willing to actually trust it, and a clear majority call it untrustworthy. That is not a contradiction. It is just the honest state of things. People are leaning on a tool they have not decided to rely on.

The more telling picture is the trend. The same research has followed the same countries since before ChatGPT landed. As use shot up, trust drifted down and worry climbed.

Figure 2

As adoption rose, trust fell and worry climbed

So familiarity does not automatically turn into comfort. And where you live changes the picture a lot. A late-2025 Edelman poll across five countries found trust in AI running from very high to very low, with the United States, where I happen to be writing this, sitting at the bottom.

Figure 3

Trust in AI, by country

The detail I keep coming back to is tucked inside the KPMG numbers. Roughly two-thirds of people say they lean on AI output without checking it against anything at all. That is judgment handed off casually, by the same people who, a few questions earlier, said they do not fully trust the thing they are handing it to.

The uncomfortable part

So does living with this stuff every day sharpen your judgment or dull it? The honest answer the evidence gives is that it depends on what you walk in with. The flip side of that Microsoft finding is that people confident in their own ability thought more, not less. The least-quoted result in the MIT preprint is that the people who built the skill first and then picked up the tool stayed sharp. The tool seems to amplify whatever disposition you bring to it. Weak judgment gets weaker. Judgment that was already strong can get a longer reach.

That is not a comforting conclusion, especially if you were quietly hoping the machine would supply the judgment for you. It will not. It supplies fluency, and fluency is genuinely worth having. But the part that decides what matters, that holds onto the right target, that feels the small wrongness in an answer that is otherwise perfectly plausible, all of that stays with the person. It was never in the text for the model to learn.

Working next to the machine every day is what made that plain to me. The imitation has finally gotten good enough to show me the original by contrast, and I think that is the most useful thing it does. Not the drafts. The clarity about which part of the work was mine all along.

Sources

1. Ackoff, R. L. "From Data to Wisdom." Journal of Applied Systems Analysis, 16 (1989).
2. Eliot, T. S. Choruses from "The Rock" (1934).
3. Aristotle. Nicomachean Ethics, Book VI, on practical wisdom.
4. Dreyfus, H. L. What Computers Still Can't Do. MIT Press, 1992.
5. Polanyi, M. The Tacit Dimension. 1966.
6. Sparrow, B., Liu, J., and Wegner, D. M. "Google Effects on Memory." Science, 333 (2011).
7. Dahmani, L., and Bohbot, V. D. "Habitual use of GPS negatively impacts spatial memory." Scientific Reports, 10 (2020).
8. Lee, H.-P., et al. "The Impact of Generative AI on Critical Thinking." Proceedings of CHI 2025. Peer reviewed. n = 319.
9. Kosmyna, N., et al. "Your Brain on ChatGPT." arXiv preprint, 2025. Not peer reviewed; treat as preliminary.
10. Plato. Phaedrus, 274c to 275b.
11. Squarciafico, H. Venice, 1477, on the abundance of books.
12. Thoreau, H. D. Walden, 1854.
13. Gillespie, N., et al. Trust, Attitudes and Use of AI: A Global Study 2025. University of Melbourne and KPMG.
14. Edelman. Trust Barometer Flash Poll: Trust and AI at a Crossroads, Fall 2025.