Handing everyday thinking to generative artificial intelligence (AI) tools has been shown to weaken memory, decision-making and persistence, as chatbots settle into classrooms, offices and courtrooms.
A single plain-language prompt is now enough to draw a usable answer out of services such as ChatGPT or Claude, which will draft emails, write code or plan trips.
A study of 1,222 people by researchers at Carnegie Mellon University, Oxford, the Massachusetts Institute of Technology (MIT) and the University of California, Los Angeles found AI help lifted performance in the short term but left users worse off once the tools were taken away.
Across three randomised trials on arithmetic and reading comprehension, participants who had leaned on AI later performed significantly worse unaided and gave up sooner, the paper, still under peer review, said.
Persistence underpins skill acquisition and is among the strongest predictors of long-term learning, the authors wrote.
Grace Liu, the Carnegie Mellon doctoral student who led the work, said the speed of AI answers “removes learning opportunities” from users.
She said what set the technology apart from earlier aids was its reach across almost any reasoning task. Electronic calculators solved equations though left the method in human hands.
A separate MIT Media Lab study published in June 2025 fitted EEG headsets to 54 people writing essays with ChatGPT, a search engine or no tools.
The chatbot group showed the weakest brain connectivity and struggled to quote back sentences they had just written, an effect the authors labelled “cognitive debt”.
“Human beings have a strong tendency to save energy,” said Johann Chevalère, a social and cognitive psychology researcher at France’s publicly funded CNRS.
He said people routinely reached for shortcuts rather than paying the cognitive cost of studying information in depth, and that generative AI could deepen the habit. A brain that never performed an activity would not maintain the connections behind it, he said.
Developers have started building “Socratic” modes that offer hints rather than answers, among them the study mode in OpenAI’s ChatGPT and guided learning in Google’s Gemini.
Microsoft told AFP it had built warnings about mistakes into Copilot and prompted users to check its answers. “The risk of excessive cognitive offloading is real,” the company said, adding that users had to be trained properly.
Brussels has meanwhile been loosening its own answer. Article 4 of the European Union’s AI Act has required firms to ensure staff are AI-literate since February 2, 2025, though the European Commission’s Digital Omnibus would shift that duty on to member states.
No large-scale, long-term evidence yet exists of what the technology does to human brains, leaving schools, employers and regulators to act on early findings.
