On this page

Synthesis: Kosmyna et al. (2025) use electroencephalography (EEG) to probe the neural and behavioral consequences of LLM-assisted essay writing. Across three sessions, 54 participants wrote essays under three conditions — Large Language Models (LLMs) (ChatGPT), Search Engine, and Brain-only (no tools) — then in a fourth session a crossover reassigned LLM users to no-tools (LLM-to-Brain) and Brain-only users to LLM (Brain-to-LLM). EEG revealed that brain connectivity systematically scaled down with the amount of external support: Brain-only participants showed the strongest, most distributed neural networks; Search Engine users moderate engagement; and LLM users the weakest connectivity. Over four months, LLM users consistently underperformed at neural, linguistic, and behavioral levels — the study frames this as an accumulation of cognitive debt. LLM users reported the lowest essay ownership and struggled to accurately quote their own work.

Key Findings

  • Three-group EEG design. 54 participants (18 completed the crossover session) wrote essays using an LLM (ChatGPT), a search engine, or no tool, across three sessions, followed by a fourth crossover session (LLM-to-Brain, Brain-to-LLM). EEG assessed cognitive engagement/load; essays were analyzed with NLP and scored by human teachers and an AI judge.
  • Cognitive activity scales down with external support. EEG showed significantly different neural connectivity across groups: Brain-only had the strongest, most distributed networks; Search Engine users showed moderate engagement; LLM users displayed the weakest connectivity.
  • Session-4 crossover reveals under-engagement. LLM-to-Brain participants showed reduced alpha and beta connectivity (under-engagement) when the tool was removed; Brain-to-LLM users showed higher memory recall and activation of occipito-parietal and prefrontal areas, similar to search-engine users.
  • Ownership is lowest with LLMs. Self-reported essay ownership was lowest in the LLM group and highest in the Brain-only group. LLM users also struggled to accurately quote their own work.
  • Consistent homogeneity within groups. Named-entity recognition (NERs), n-grams, and topic ontology showed within-group homogeneity, while between-group neural and linguistic differences were robust.
  • Cognitive debt accumulation. Over four months, LLM users underperformed at neural, linguistic, and behavioral levels, raising concerns about the long-term educational implications of LLM reliance.

What this means for practice

  • Instructors. Keep at least some writing tasks tool-free: in the 54-participant EEG study, the Brain-only group showed the strongest and most distributed neural connectivity and the highest essay ownership, while LLM users showed the weakest connectivity.
  • Instructors. Require students to work with their AI-assisted text after the tool is gone — quote it, revise it, or explain its argument — because LLM users in this study struggled to quote their own essays accurately, and the LLM-to-Brain crossover showed reduced alpha and beta connectivity when support was removed.
  • Instructors. Treat fluency with the tool as a warning sign rather than a success criterion: convenience in the moment came with lower engagement and, over four months, neural, linguistic, and behavioral underperformance.
  • Designers. Build writing tools that preserve ownership and self-monitoring rather than replace them, since self-reported essay ownership was lowest in the LLM condition and the study connects that detachment to authorship concerns.
  • Researchers. Measure retention and transfer separately from in-task performance, because this design shows the two can diverge in opposite directions.

Limitations

  • 54 participants (aged 18–39, M = 22.9) recruited from five universities in the greater Boston area — MIT, Wellesley, Harvard, Tufts, and Northeastern — an unusually selective, English-speaking sample; 60 were originally recruited and 55 completed the full protocol.
  • Only 18 of the 54 participants attended Session 4, the crossover that carries the study's central claim, because it was optional and depended on scheduling.
  • Essays were written under a 20-minute time limit with 32-channel EEG hardware, so the writing task and context are not typical of coursework, and the study reports a preprint under review rather than a peer-reviewed article.
  • The central outcomes are proxies: EEG connectivity differences and participants' self-reported ownership stand in for learning and engagement, and no delayed test of essay knowledge or writing skill was reported.

Connected Concepts

Connected Articles

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.

Embed this page

Copy the code below to embed a chromeless version of this page in a learning management system or other website. The embedded view hides the site header, navigation, and footer.