MIT Study Shows Neurological Effects of ChatGPT in Education
A study by MIT Media Lab led by French neurotechnology researcher Nataliya Kosmyna provides the first measurable evidence of the neurological effects of ChatGPT use in writing. The four-month study shows that the use of Large Language Models in educational contexts can lead to measurable changes in brain activity and cognitive abilities.
Study Design and Methodology
The researchers divided 54 participants aged 18 to 39 from the greater Boston area into three groups that were to write SAT essays over several months. The first group used ChatGPT (GPT-4o), the second group used conventional search engines like Google, while the third group wrote without any aids. Using high-resolution electroencephalography (EEG), the scientists monitored brain activity in 32 different brain regions throughout the entire writing process.
In a fourth session, the groups switched their methods: those who had previously used ChatGPT had to write without aids, while the "brain-only" group was now allowed to use ChatGPT. This role reversal proved particularly revealing for the long-term effects of AI use.
Systematic Decline in Neural Connectivity
The EEG analyses showed clear differences in neural connectivity between the three groups. Brain connectivity systematically decreased the more external support was used. While the "brain-only" group showed the strongest and most extensive neural networks, the search engine group showed moderate cognitive activation. The ChatGPT users, however, showed the weakest overall coupling, with up to 55 percent lower cognitive activation than the group without aids.
Particularly affected were the alpha and beta frequency bands, which are central to memory and language processing. The ChatGPT group also showed reduced frontal-medial theta activity, which is associated with working memory demands.
The Concept of Cognitive Debt
At the center of the study is the new concept of "cognitive debt". Participants who used ChatGPT for four months continued to show weaker alpha and beta networks, as well as significantly reduced memory ability, even when writing later without AI support. While 83 percent of participants from the "brain-only" group could quote from their own essays in the first session, only 17 percent of ChatGPT users managed to do so.
The researchers also found that the ChatGPT group developed a low sense of ownership for their texts. Many participants reported a "fragmented and contradictory sense of authorship," with some claiming full ownership, others explicitly denying it, and many seeing themselves only partially as authors.
Risks for Education
Early AI Introduction as Developmental Risk
Nataliya Kosmyna particularly emphasizes the danger of premature introduction of AI tools in educational institutions. "What really motivated me to publish the results now, before we wait for a complete peer review, is the fear that in six to eight months a politician might decide: 'Let's introduce GPT in kindergarten.' I think that would be absolutely bad and harmful," she explains. In her view, "developing brains are at the highest risk".
Automation Instead of Learning
Already in the third session, many ChatGPT participants primarily copied text from the AI with only minimal editing. This apparent "efficiency," however, comes at the expense of memory integration and deeper understanding. As Kosmyna explains: "The task was completed, and you could say it was efficient and convenient. But as we show in the study, you have integrated practically nothing of it into your memory networks".
Deterioration on Multiple Levels
The ChatGPT group consistently performed worse than the "brain-only" group over the four months at neural, linguistic, and behavioral levels. The essays of the AI-assisted group, evaluated by human teachers, were characterized by "near-perfect language and structure use while simultaneously failing to provide personal insights or clear statements".
Promising Findings on Proper AI Use
Not all results of the study are pessimistic. The "brain-to-LLM" group, participants who first wrote without aids and then used ChatGPT, showed remarkable improvements. This group showed higher memory performance and demonstrated a significant increase in brain connectivity across all EEG frequency bands.
The search engine group also maintained high cognitive activity and satisfaction. This underscores the importance of tool selection and shows that not all digital aids have the same effects.
Implications for Educational Policy and Teaching Practice
Sequence is Crucial
The study impressively demonstrates that the sequence of AI introduction is of central importance. As the researchers write: "From an educational policy perspective, these findings suggest that strategic timing of AI tool introduction after initial self-directed effort can improve engagement and neural integration".
Adaptation of Educational Standards Required
Similar to the introduction of calculators in the 1970s, new educational standards must also be developed for AI tools. Instead of solving the same tasks with AI help, more complex challenges should be posed that integrate AI as a tool, but not as a replacement for critical thinking.
Warning Against Excessive Dependence
The study warns against excessive dependence on AI tools, which can lead to what researchers call "metacognitive laziness". In this state, learners completely delegate critical thinking processes to AI and thereby lose the ability to solve problems independently.
Study Limitations and Future Research
The researchers emphasize various important limitations of their investigation. The limited number of participants from a specific geographical region, the exclusive focus on ChatGPT without considering other Large Language Models, and the context-specific results that only concern essay writing limit the generalizability of the findings.
The study was published as a preprint in June 2025 and is still awaiting peer review. Nevertheless, the preliminary results underscore the urgency of understanding the cognitive effects of AI in education before making far-reaching educational policy decisions.
Conclusion and Outlook
The MIT study provides important first evidence that the type and timing of AI introduction in education are crucial for their effects on cognitive development. While AI tools can indeed improve learning, this only happens when they are used as a complement to human cognition and not as a replacement. The results suggest that learners should first develop basic cognitive abilities without AI support before strategically using these tools to improve their work.
For educational institutions, this means that a thoughtful integration strategy is required that considers both the potential and risks of AI tools. The study shows that AI can indeed be valuable, but only when used in the right context and at the right time.
Sources:
