Researchers at the MIT Media Lab, in collaboration with researchers from Wellesley College and the Massachusetts College of Art and Design, wanted to find out if using ChatGPT changes how people think and learn when writing. Findings, which were first published in a pre-print article in arXiv on 10 June 2025, have raised concerns about the long-term educational implications of using large language models or LLMs and underscored the need to investigate further the role and impact of artificial intelligence in learning.
Using ChatGPT or Other LLM-Based Writing Assistant: An MIT Study Found Accumulation of Cognitive Debt When Using an AI Assistant to Write Essay
Background
Writing assignments in the academe promotes various cognitive and linguistic abilities essential to intellectual growth. These include critical thinking and reasoning, memory formation and retrieval, language and communication skills, executive function, creativity and imagination, emotional processing and self-awareness, and cognitive load management.
The eight-member team of researchers, which included corresponding author Nataliya Kosmyna and senior author Pattie Maes, both from MIT Media Lab, recruited 54 participants between the ages of 18 and 39 and divided them into three groups for the first three sessions. These were the LLM group, the search engine group, and the so-called brain-only group.
All groups were asked to write different essays based on SAT topics. There were 3 topics for each session for a total of 9 topics throughout the three sessions. Each group was also given different mechanics. The LLM group used ChatGPT to substantiate their writing. The other group used websites. The brain-only group wrote the essays without any tools.
It is worth mentioning that 35 participants reported pursuing undergraduate studies and another 14 participants were enrolled in postgraduate studies. Nevertheless, in the fourth session, only a total of 18 participants remained. This session switched things up. The LLM group was asked to write an essay without tools and the brain-only group wrote an essay using an LLM.
Nevertheless, to measure impact, the researchers used a variety of methods. They used an EEG to record brain activity across 32 regions. Each essay underwent a natural language processing or NLP analysis to look for patterns. The researchers also conducted interviews with the participants after each session. Essays were evaluated by human teachers and AI.
Findings
• Distinctions in Brain Activity and Cognitive Strategies: EEG results showed that each group used significantly different brain connectivity patterns. This indicated that they were employing different mental strategies to write. The brain-only group had the strongest brain network engagement while the search engine group showed moderate brain engagement. The LLM group had the weakest neural activity.
• Patterns in the Content and Structure of the Essays: Essays from each group were consistently different in their structure, phrasing, and topics. ChatGPT users or members of the LLM group produced written outputs that were similar to each other in both form and substance. This indicated possible homogenization.
• Perceived Sense of Ownership of the Written Outputs: The brain-only group felt the most ownership and also remembered more from their essays and the search engine group had a moderate sense of ownership. The LLM group felt less connected to their writing. This group also struggled to quote from their written outputs. This was evident even a few minutes after the essays were written and submitted.
• Specific Results of Switching Tools in the Fourth Session: Weaker neural connectivity and under-engagement in certain brain areas were observed in the LLM group who switched to the brain-only task. The brain-only group, which then used LLM, had higher memory recall and re-engagement of brain areas associated with visual processing.
• Distinctive Impact on Level of Learning Skills Over Time: The findings suggest a pressing matter of a likely decrease in learning skills with LLM use. Specifically, over the course of the four-month participant engagement, members of the LLM group consistently performed worse than the participants in the brain-only group at tests for cognitive abilities as determined by neural, linguistic, and scoring levels.
Implications
The findings above and the overall study provide preliminary evidence that long-term use of LLMs in educational contexts could lead to a measurable decrease in learning skills despite offering some benefits. These findings highlight the importance of understanding the cognitive and practical implications of AI on how individuals, especially students, learn.
Research pursuits aimed at assessing the impacts of artificial intelligence, especially generative artificial intelligence or Gen AI tools like chatbots and reasoning models, are ongoing. An earlier four-study research published in April 2025 and involving 3500 participants revealed that Gen AI provides an immediate performance boost at the expense of intrinsic motivation.
Another study by researchers at Microsoft and Carnegie Mellon University, which was published in the second quarter of 2025 and involved surveying 319 individuals in professions that use Gen AI at least once a week, found that generative AI tools improved efficiency but inhibited critical thinking and promoted long-term overreliance on the technology.
Findings from the study of Wharton professors L. Meincke and G. Nave and Wharton researcher C. Terwiesch showed that individuals who used ChatGPT came up with analogous concepts and demonstrated reduced overall idea diversity. Individuals who used their own thoughts and web searches came up with a more diversified or broader range of creative ideas.
The aforementioned studies raise important questions about the cognitive cost and implications of convenience. These suggest that generative AI and models like LLMs and reasoning models can be useful tools but come with drawbacks. Individuals and organizations must be cautious about how these tools or applications are integrated into learning and productivity.
FURTHER READINGS AND REFERENCES
- Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., and Maes, P. 2025. “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task (Version 1).” arXiv. DOI: 48550/ARXIV.2506.08872
- Lee, H.-P. Hank, Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., and Wilson, N. 2025. “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.” In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (pp. 1–22). CHI 2025: CHI Conference on Human Factors in Computing Systems. ACM. DOI: 1145/3706598.3713778
- Meincke, L., Nave, G., and Terwiesch, C. 2025. “ChatGPT Decreases Idea Diversity in Brainstorming.” Nature Human Behaviour. DOI: 1038/s41562-025-02173-x
- Wu, S., Liu, Y., Ruan, M., Chen, S., and Xie, X.-Y. 2025. “Human-Generative AI Collaboration Enhances Task Performance but Undermines Human’s Intrinsic Motivation.” Scientific Reports. 15(1). DOI: 1038/s41598-025-98385-2