Short answer: it can, and the mechanism is well understood. It depends entirely on which part of the work you hand over.
The design that matters
Most claims about AI and learning rest on the wrong measurement. Asking whether students perform better while using a chatbot answers a question nobody needs answered, because the exam will not have one running.
The informative studies are the ones that take the tool away and test afterwards. Those studies exist now, and they point in a consistent direction.
What they find
In a large trial with high school mathematics students, a group given access to an ordinary chatbot interface outperformed the control group while they had it, and underperformed the control group once it was removed. Practice performance went up. Learning went down. Those are not the same variable, and this is the clearest demonstration of that anyone has produced.
Smaller experiments agree. In one, students who used a chatbot to complete a written assignment scored lower on an unannounced test six weeks later than students who did the same assignment without one. Diary studies comparing chatbot use against ordinary web search have found worse outcomes for the chatbot group, with the difference concentrated in the kind of understanding that requires connecting ideas rather than retrieving facts.
None of this is unique to AI, which is why the finding was predictable. Sparrow, Liu and Wegner showed in 2011 that information people know is stored elsewhere is remembered less well, even when retrieving it is trivially easy. Risko and Gilbert’s work on cognitive offloading documented the general rule: when asking is easier than thinking, people ask. A chatbot is the most frictionless version of that arrangement ever built.
The finding that should worry you most
Several studies report that people using these tools rate their own understanding higher while performing worse.
Falling ability alongside rising confidence is the single most dangerous pattern in studying, with or without a computer involved, because it removes the signal that would have told you to do something differently. You do not investigate a problem you do not believe you have. It is the same failure as the fluency trap, arriving faster and with better prose.
The rule that follows
Never let it perform a retrieval you could have attempted.
The damage is not caused by the technology being present. It is caused by it performing the specific cognitive work that learning consists of — recalling, explaining, connecting, struggling briefly and then succeeding. Hand that over and there is nothing left to learn from, however good the output looks. Hand over the surrounding work and nothing is lost at all.
Uses that survive the rule
- Marking a recall attempt after you have written it, and telling you what you missed.
- Generating practice questions that you then answer closed-book.
- Arguing against an explanation you produced yourself.
- Explaining something you have already attempted and failed to work out.
- Administrative work around studying that was never going to teach you anything — formatting, scheduling, tidying references.
Uses that do not
- Summarising a chapter you have not read.
- Explaining a concept before you have attempted it yourself.
- Producing an answer you then read and recognise, which feels like understanding and is not.
- Anything that replaces a blank page you could have filled badly.
The pattern across both lists: the tool is safe when it arrives after your own effort and dangerous when it arrives instead of it. Order matters more than frequency. A student who attempts everything first and then checks with a chatbot is using it well. A student who asks first and reads the answer has replaced their education with a reading exercise.
An honest caveat
This evidence is newer and thinner than the evidence for spacing or retrieval practice. The tools change faster than studies can be run, most samples are small, and publication favours striking results in both directions. Read confident claims about AI and learning with suspicion, including the ones on this page.
What I would defend is the mechanism rather than any particular effect size. Offloading the cognitive work of learning has produced the same result in every form it has taken for fifty years, and there is no reason this form would be the exception.
In the book
Chapter 23 treats this at length, including what to do when a course requires you to use these tools. Appendix B lists the studies in full.