AI promises to save us the trouble of thinking. That is useful right up to the point when thinking is the thing we were supposed to be learning.
The concern is not that every chatbot session causes damage. It is that making AI a routine substitute for effort could weaken the habits and institutions through which people develop knowledge. To see why, start with a familiar but incomplete picture of the mind: the brain as a computer.
The brain is not just an input-output machine
The computational model treats thought as a sequence. We take in information, process it, then act. It has helped technologists build increasingly capable machines, from basic calculators to generative AI. But as a model of human thought, it leaves a lot out.
Neuroscientist Paul Cisek proposes a different starting point. Rather than treating the brain chiefly as an information processor, his account sees it as part of a feedback-control system. We act on our surroundings, those actions change what we sense, and our next actions depend on that altered situation. Perception and movement are not separate stages on an assembly line.
Consider an outfielder catching a fly ball. A strict computational account suggests the player must estimate the ball's speed, calculate gravity's effect and work out where it will land. The feedback account offers a simpler description: keep the ball in roughly the same place in your field of vision while moving to stay beneath it.
The second explanation better matches the experience of playing catch, and it does not require the brain to solve a physics problem consciously or unconsciously before the body can move. The body is part of the thinking process, not merely the vehicle that carries out its result.
Why it matters: If minds are understood only as machines that take in information and produce answers, then AI looks like a natural learning aid. If minds are also systems that develop through action, adjustment and experience, replacing those activities may come at a cost.
Cisek's account connects this feedback idea to the history of nervous systems. As animals encountered new ways to move through and interact with their environments, new behaviors and capacities emerged. The brain, on this view, developed alongside the expanding range of things an organism could do.
That history does not fit neatly into a simple input-computation-output diagram. It suggests that human cognition is bound up with controlling our interactions with the world, not just processing descriptions of it.
Learning is a social feedback loop
The environment that shapes human cognition is not only physical. It is social, too.
People learned by watching and copying one another. Shared gestures and practices made it possible to pass on useful skills, coordinate activities and build culture. Imitating sounds helped give rise to oral language, while writing allowed ideas to travel beyond a speaker's presence and across generations. Communities developed education and institutions for sharing knowledge and making collective decisions.
These practices are not decorative additions to an otherwise solitary mind. They are part of how people learn. A lesson, a conversation or an argument gives someone a chance to test an idea, notice what they missed and revise their understanding. The process can be slow. That is partly the point.
Generative AI changes the terms of that exchange. Large language models draw on information produced by people and can return a plausible response without requiring the user to reconstruct the reasoning behind it. The technology can therefore make accumulated human knowledge easier to access while also making it easier to skip the work of acquiring knowledge for oneself.
The distinction is easy to lose because the output can look like learning. A student asks a question, receives a clear explanation and moves on. But receiving an answer is not the same as being able to remember it, assess it or use it without help.
The cognitive hot dog
The critique compares routine AI use to a hot dog. The analogy is not that a single hot dog, or a single AI prompt, will ruin anyone's health. The warning is about what happens when something convenient becomes a regular substitute for what a person needs to develop or maintain.
Human bodies evolved in conditions where calorie-dense food was scarce. Modern food systems made it abundant, which changed the consequences of seeking it out. In a similar way, AI can provide quick cognitive relief in situations where effortful thought is uncomfortable. But the habit of choosing the easy answer may interfere with building the skills that make independent judgment possible.
François Chollet, formerly a software engineer at Google, has described AI as cognitive automation: software that represents human abstractions and performs tasks people would otherwise do themselves. The term gets at both the appeal and the risk. Automation can save effort, but some forms of effort are also how people learn.
That makes the comparison with earlier technologies imperfect. Calculators can take over arithmetic, and written language can store information outside a person's memory. Generative AI is also being promoted for tasks that involve forming explanations, ideas and judgments. Its use may therefore affect not just how people retrieve facts, but how often they practice the thinking those tasks demand.
Schools are the clearest test
Education brings the trade-off into focus. Students need information, but they also need practice working through problems, checking claims and forming their own explanations. A tool that makes assignments easier to complete can make it harder to tell whether learning took place.
The warning is not merely theoretical. A study in China described in the report found that thousands of students largely stopped doing homework after they began using AI, with a substantial cost to their learning. Other research cited there suggests that students using large language models for learning may become more accustomed to relaxing their judgment and critical thinking in other settings.
Those findings do not establish that every use of AI harms every student. They do, however, challenge the reassuring idea that a chatbot becomes a tutor simply because it can explain things. A tutor can ask a learner to try, spot a misunderstanding and require another attempt. A system that supplies finished work may do the opposite, unless its use is designed around the student's effort rather than the machine's fluency.
The commercial enthusiasm is not subtle. In a July 2025 podcast about AI in education, OpenAI vice president of education Leah Belsky said learners made up more than half of ChatGPT's 900 million average monthly users and described the service as the world's largest learning platform. Anthropic has also introduced Claude for Teachers, offering practicing classroom educators free access to premium models.
Those offerings may have legitimate uses. But counting users does not show that students are learning, any more than counting visits to a library proves that anyone read the books. The harder question is whether the tools support the work of learning or quietly replace it.
Keep some thinking deliberately unaided
One proposed response is cognitive immunization: protect the practices that keep people reasoning for themselves. The prescription is practical rather than mystical. Make room for solving some problems without AI, checking answers against other evidence, discussing disagreements and spending time away from the tools.
For schools, that means deciding which tasks benefit from AI and which should remain a student's own work. It also means treating access and convenience as separate questions from educational value. A chatbot can be available everywhere and still be the wrong answer to a particular teaching problem.
Some authorities and school groups are already drawing limits. Norway has banned essentially all uses of AI in school for students under 13. In the United States, a major teachers' union has called for chatbots to be prohibited in elementary schools, while school districts in Los Angeles and New York City have implemented bans.
Restrictions carry trade-offs. A blanket rule may prevent helpful uses along with harmful ones, and a ban alone cannot teach students how to assess AI-generated claims. But allowing unrestricted use is a choice, too. It hands a powerful role in the learning process to products whose convenience is easy to measure and whose effects on durable knowledge are harder to see.
A more careful approach would preserve room for students to think, make mistakes and explain their work, while using AI where it supports those activities rather than displacing them. That is less tidy than announcing that every classroom needs a chatbot. It is also closer to the actual job of education.
The argument reaches beyond schools. If people routinely delegate judgment, the social conditions that support independent reasoning may weaken, making delegation still more tempting. The cycle is familiar: use a tool because it is easy, lose some practice doing the task yourself, then find the tool even harder to put down.
Human cognition was shaped by millions of years of biological development and by generations of shared learning. AI may be useful within that history, but usefulness is not the same as harmlessness. The question is not whether we can get machines to think for us. It is how much thinking we are willing to stop doing ourselves.
Sources
- Our minds aren’t equipped to handle AI — The Verge