By: Tracey Tokuhama-Espinosa, Ph.D.
Harvard University Extension School
5 June 2026
Science Counts: a commentary on a published scientific article that brings some new data on Science of Learning
AI and Education is a topic that invites significant controversy and a broader debate on the goals of education. AI can be seen as a savior for poor schools with few resources, and a villain when seen as stripping away basic cognitive skills that go unrehearsed when offloaded to a device like a phone or computer. Since coming on the scene in November 2022, ChatGPT has led the vibrant exchange about AI’s disruptive challenges to basic education.
In the Summer of 2025, an article was published that quickly became a part of many conversations. Kosmyna and colleagues at MIT wrote an article called, “Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task.” They use EEG data to review 54 participants divided into three groups as they wrote an essay: (a) using LLMs like ChatGPT, (b) those who used a regular search engine like Google, and (c) those who used their brains alone. Group (c) used far more neural networks than the search engine group (b), and the weakest coupling came from the LLM ChatGPT group (a).
While there were benefits to personalized and tailored content in the LLM group, there appeared to be far more negative outcomes than positive. First and foremost, the researchers identified “diminished critical thinking capacity,” “decreased deep analytical thinking,” and possible “cognitive atrophy” due to reduced use of established neural networks. This suggests that without “using it” the students “lost it.” Lack of stimulation of key neural networks due to offloading tasks to the LLM reduced participants’ long-term abilities to rely on those networks in the future.
Second, and equally as alarming was the “diminished prospects for independent problem solving,” as students became accustomed to ChatGPT offering a single, right answer, rather than allowing learners to struggle with a question and doubt premises as they explored possible angles. There appeared to be no place for “lateral thinking and independent judgement” when using ChatGPT. This contrasted with the brain-only group who had to review past memories and information, creatively combine old and new concepts, entertain multiple possible approaches, rephrase and reframe ideas and search for appropriate vocabulary and formatting to construct coherent arguments.
While the debate continues, it is clear that at least three ideas have surfaced as meaningful contributions to the learning sciences thanks to this article and the debates around ChatGPT and the brain. First, the brain adapts to what it does most. Without rehearsal, core neural networks will go unrehearsed. Second, “use it or lose it.” If those networks go unrehearsed, they will possibly atrophy and no longer be available when needed. Third, this experiment combined with other new evidence about the neuroscience of writing (Tokuhama-Espinosa et al., 2024) shows that “Writing is thinking and thinking is writing. People should write to think, and think to write.”
It is true that ChatGPT can write for humans and often do so much better than most when prompted well. It is also true that many people suffer when it comes to writing and feel that their efforts are not worth the outcome. Kosmyna and colleagues have shown, however, that the struggle is what keeps the neural networks in shape, and allow deeper learning to be built on top of foundational networks: “no pain, no gain”. If critical thinking, deeper learning, lateral thinking and independent judgement are desired, and avoiding cognitive atrophy is valued, then the effort is worth it. Teaching writing, the old fashioned way, may be the key to improved education in the age of AI.
Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X. H., Beresnitzky, A. V., … & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv preprint arXiv:2506.08872, 4.
