Does using AI make you worse at thinking?

The worry is ancient. In Plato’s Phaedrus the Egyptian king Thamus is offered the invention of writing and refuses to be impressed: it will, he says, produce forgetfulness in those who learn it, and give them the appearance of wisdom without the reality. Writing went on to make most of human learning possible. That history is a reason to be slow to panic, and also a reason to look at what was in fact lost and kept.

What the evidence shows

Psychologists call the general practice cognitive offloading: using something outside the head to reduce the work inside it. Its effects on memory are well established. Betsy Sparrow and colleagues found that people who expect to be able to look something up remember the fact less and where to find it more.

On AI specifically the evidence is recent and should be read with care. Michael Gerlich surveyed several hundred people and found that heavier use of AI tools went with lower scores on measures of critical thinking, with offloading accounting for much of the link. A study by researchers at Microsoft and Carnegie Mellon of knowledge workers found that the more confidence people had in the AI, the less critical effort they reported putting in, while confidence in their own ability went with more.

These are associations, largely self-reported, and they do not show which way the influence runs; they are, though, what one would predict from the plain fact that a skill left unused declines.

Why AI is a different kind of tool

A notebook stores what you put in it. A calculator carries out an operation you chose. In both cases you did the thinking and the tool held or executed part of it. Andy Clark and David Chalmers went so far as to argue that such tools, reliably used, are part of the mind.

A language model can be asked for the argument itself, the plan, the summary of a text you have not read. What is handed over is then the work of understanding. And here a second problem appears that the notebook never raised: to notice that the system has gone wrong, you need the very competence you did not exercise.

Where I stand

I do not think the right question is whether to rely on these tools. We rely on instruments and on other people for most of what we know, and could not do otherwise. The right question is what has to stay with the person for the result to be theirs.

My account of how one can come to know something through an AI system is that the user is reading an instrument: the warrant comes from the user’s own competent use, their sense of what the system is good at, where it fails and how to check it. That competence is the one thing that cannot be delegated to the tool, because it is what makes relying on the tool reasonable.

The test I would apply to any use is whether, after the system has done its part, you could tell if it had done it badly. If you could, you have saved effort. If you could not, you have a result you are in no position to trust, however good it looks, and each repetition makes the position a little worse.