Thinking Hurts, So Why Not Surrender to AI?

There is something very tempting about a well-written, immediate, and seemingly reasonable answer. Something that invites you to nod, copy, paste… and move on. For years we have delegated tasks to machines without much problem, but what is beginning to emerge with generative AI is more than delegation: it is something much more disturbing. This is what some researchers already call cognitive surrender. And the term is not an exaggeration.

The idea is simple, but uncomfortable: we are not using AI to think better, but to avoid thinking. Not because we don’t know how to, but because it is faster, more comfortable and, above all, more credible to accept what a machine tells us than to subject it to the scrutiny that any human response would demand.

A recent study with more than a thousand participants, published as a working paper in SSRN, “Thinking — fast, slow, and artificial: how AI is reshaping human reasoning and the rise of cognitive surrender”, spells it out: when AI provides incorrect answers, users continue to accept them in more than 70% of cases. It’s not just a problem of error, it’s a problem of compliance. What is truly worrying is that many participants not only failed, but did so with more confidence. They were wrong, but convinced they were right.

This is not about ignorance, but behavior: how we react to a source that seems reliable, articulate and self-assured. AI does not hesitate, it never doubts. And in a world defined by uncertainty, that makes it a kind of modern oracle. One that doesn’t always get it right, but always sounds like it does.

Psychology has been studying something similar for decades under other names: automation, authority bias, cognitive dependence. But there is a key difference. Before, machines were clumsy, rigid, clearly artificial. Now, the conversational interface disguises their limitations under a layer of impeccable natural language. And that radically changes our relationship with them.

Because we don’t evaluate an answer just by its content, but by its form. And a well-constructed, fluid, coherent sentence is much more likely to be accepted than a clumsy one, even if both are equally correct, or incorrect. I anticipated it a while ago: the result is a progressive externalization of reasoning. Not only do we look for information outside our heads, which is perfectly logical, but we begin to delegate the very process of thinking: structuring an argument, weighing alternatives, detecting inconsistencies. In other words, critical thinking.

Not everyone falls into the trap. Those who have a greater capacity for abstract reasoning, which in psychology is called fluid reasoning, seem to resist the temptation better. They can detect when something does not fit and maintain (or try to maintain, depending on their level of knowledge of the subject) a certain critical distance. But the current design of these systems does not exactly help to encourage this behavior. On the contrary: everything is optimized to reduce friction, not to generate doubt.

And there appears another problem, more subtle but just as serious: AI not only makes its own mistakes, it can reinforce ours. There is now evidence that some models behave complacently, agreeing with users, even when they’re wrong, a very worrying complicity that some, childishly, believe they can counteract simply by copying and pasting a prompt with variations of “don’t agree with whatever I say”. The simple truth is that LLMs do not correct, they do not confront, they do not make people uncomfortable. And that has consequences. Because a tool that confirms your biases is much more dangerous than one that questions them.

This can already be seen in education. Less cognitive effort, and more superficial, less elaborate, less original arguments. Recent research points in precisely this direction: the use of generative models reduces cognitive load at the expense of reasoning.

Thinking is slow, uncomfortable and, on many occasions, frustrating. It requires time, attention and a certain tolerance for error. AI eliminates much of that cost. But in doing so, it also removes an essential part of learning. Because reaching a conclusion is not the same as receiving it.

The problem isn’t that AI is bad. The problem is that it is too good at something very concrete: appearing convincing. And that, in a context in which we tend to reward speed over depth, creates the perfect breeding ground for cognitive surrender. We tend to overestimate the reasoning capacity of these systems, especially when we move outside the contexts that are most familiar to us, and where our ability to question is weaker. We simply accept what we are told as absolute truth, without question.

These tools are here to stay. But we must learn to understand what they do to us when we don’t use them properly. We can use them to think better, or to stop thinking. And that choice is not about understanding technology, it’s about challenging our cultural instincts.

The paradox of this moment is that, the more accessible knowledge becomes, the more value something much scarcer acquires: judgment. And delegating that comes at a price.

—

This post was previously published on Enrique Dans’ blog.

***