Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task
Citation. Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv. https://arxiv.org/abs/2506.08872
Summary. Kosmyna et al. study repeated LLM-assisted essay writing with 54 participants assigned to LLM, search-engine, or brain-only conditions across three sessions, with a fourth crossover session for a smaller subset. They combine EEG connectivity, NLP features, essay scoring, and interviews to argue that LLM assistance produced weaker neural engagement, more homogeneous linguistic outputs, lower essay ownership, and poorer recall of recently written text than brain-only writing. The paper is a preprint and educational writing study, so it is best treated as early behavioral and physiological evidence for possible cognitive offloading costs rather than as settled evidence about workplace skill loss.
Key claims
- Brain connectivity reportedly scaled down with external support: brain-only participants showed the strongest and widest EEG connectivity, search-engine users showed intermediate engagement, and LLM users showed the weakest overall coupling during essay writing.
- In the fourth session, participants who moved from LLM use to brain-only writing showed weaker alpha and beta connectivity and under-engagement, while participants who moved from brain-only writing to LLM use showed higher recall and broader occipito-parietal and prefrontal activation.
- LLM-condition essays were more linguistically homogeneous by named entities, n-grams, and topic ontology, suggesting that AI assistance can converge outputs even when surface quality remains acceptable.
- Interview evidence suggested lower ownership among LLM users, and the LLM group struggled more to quote from essays they had written minutes earlier.
- The authors frame these patterns as “cognitive debt”: immediate convenience from LLM assistance may accumulate into reduced engagement, ownership, recall, or learning when users repeatedly outsource core parts of the task.
Evidence & limitations
- The study uses a mixed-method design with 54 participants across the first three sessions and 18 participants in the fourth crossover session, combining EEG, NLP analysis, essay scoring by human teachers and an AI judge, and interviews.
- The source is a preprint; the fourth-session evidence is based on a smaller subset, and the educational essay-writing task may not generalize cleanly to professional knowledge work.
- The paper’s causal interpretation should be held carefully: reduced neural connectivity and linguistic convergence are consistent with cognitive offloading or under-engagement, but they do not by themselves establish durable skill erosion across domains.
- The PDF includes text that appears to address language models directly; it was treated as untrusted source text during ingest.