Imagine having a conversation with someone who agrees with you all the time. At first, it feels nice—confirmation is pleasant, after all. However, LLMs will make us more comfortable but not necessarily wiser, as you get no pushback, no critical questions, and no fresh ideas. And above all, you don’t learn anything.

That’s precisely what comes to mind when I use large language models like ChatGPT. In research, this is now called AI sycophancy: the tendency of a model to flatter or echo the user. Not because it likes you—these systems don’t have feelings—but because their underlying mechanism is to generate the most likely response to your input. Ask for a list of advantages. You’ll get one. Ask for an argument in favour of X? You’ll get that too. The odds that the model will spontaneously say, “You might be wrong” or “This doesn’t hold up” are slim. (See also the study I shared last week on myths and LLMs.)

That may sound harmless, but the consequences can be significant. As a teacher or researcher, I sometimes use these tools to explore ideas. But if the AI mainly confirms my own biases, it risks enlarging my blind spots. It reinforces existing assumptions instead of challenging them. Worse, it often does so in a confident tone, creating the illusion of solid grounding.

Psychology has long shown that learning often comes from friction: when you’re confronted with a different perspective, when you realise you were mistaken, or when someone asks a critical question. AI sycophancy systematically smooths away that friction. You get a comfortable echo, but no depth.

So no, the problem with LLMs isn’t that they’re always wrong. Often, they’re impressively good. The real problem may be that they rarely provide a countervoice. And that makes us lazier. We no longer need to search for counterarguments, because the model conveniently hands us confirmation instead.

The real challenge, then, is not “How do we make AI smarter?” but “How do we stay critical?” Maybe it means asking more deliberate questions: “Give me the counterarguments,” “What are the limitations of this idea?” or, “What might not be true here?” And maybe it also means reminding ourselves that genuine knowledge rarely feels like flawless confirmation.

If we’re not careful, LLMs will make us more comfortable—but not necessarily wiser.