Your Brain on ChatGPT: EEG Study Reveals Cognitive Load of Writing with AI Tools

Overview

The study investigates the cognitive costs of using large language models (LLMs) in education, specifically for essay writing. Participants were assigned to three tool groups—LLM, Search Engine, and Brain-only—plus a no-tool control in later sessions, with 54 participants enrolled across sessions and 18 completing the final session. EEG tracked neural engagement, while NLP analyses and human/A I scoring evaluated outcomes. In four sessions, researchers observed how each external aid altered cognitive strategies, memory recall, and sense of ownership of the written work.

Key Takeaways

Results showed systematic differences in brain connectivity across groups, with the Brain-only condition exhibiting the strongest neural networks and the LLM group the weakest. Ownership and recall varied by group, and initial advantages from external tools evolved over time. The work is a preprint and not yet peer-reviewed, and the authors emphasize limitations and avenues for future research.

Overview and purpose

This article summarizes a cognitive neuroscience study titled Your Brain on ChatGPT, which investigates how different external tools affect cognitive load and writing performance in an educational setting. The researchers used electroencephalography (EEG) to monitor brain activity as participants wrote an essay while using three tools—OpenAI's ChatGPT (LLM), a standard search engine, or writing with the brain alone (no external aid). A fourth session reversed tool usage to test transfer and strategy changes. A total of 54 participants were recruited for Sessions 1–3, with 18 completing Session 4, which included tool-switch conditions. The study combined NLP analyses of the essays with human and AI-driven scoring, and post-session interviews to capture subjective ownership and strategies.

Quote 1: “Neural connectivity patterns differed across tool conditions, reflecting divergent cognitive strategies.” - Nataliya Kosmyna

Methods and design

Participants were divided into LLM, Search Engine, and Brain-only groups, each using their designated tool (or none in the brain-only condition). Four sessions were conducted over several months with the same group assignments in Sessions 1–3. In Session 4, LLM participants used no tools (LLM-to-Brain), while Brain-only participants used LLM (Brain-to-LLM). EEG metrics captured functional connectivity, while NLP analysis examined Named Entity Recognition (NER), n-grams, and topic ontologies. Essays were scored with input from human teachers and an AI judge. Interviews provided qualitative insight into perceived ownership and strategy changes.

Findings and interpretation

The study found robust homogeneity in linguistic features (NER, n-grams, topic ontology) within groups but significant differences in neural connectivity between groups, indicating distinct cognitive approaches. The Brain-only group showed the strongest, most widespread networks, while the LLM group demonstrated the weakest overall coupling. In Session 4, LLM-to-Brain participants exhibited weaker neural connectivity and reduced alpha/beta engagement, whereas Brain-to-LLM participants recalled more memories and re-engaged occipito-parietal and prefrontal networks, suggesting a shift toward visual processing patterns similar to the Search Engine group. Subjective ownership of essays followed a parallel pattern, with the LLM group reporting lower ownership and reduced ability to quote their own work shortly after writing.

Quote 2: “The Brain-only group exhibited the strongest, widest-ranging networks.” - Nataliya Kosmyna

Implications and limitations

The authors argue that AI-enabled writing tools can alter learning trajectories, potentially diminishing certain cognitive skills over time if reliance on external aids persists. They emphasize that benefits appeared early but performance across neural, linguistic, and scoring measures lagged behind Brain-only writing across sessions. The study is a preprint on ArXiv (as of June 2025) and caution is advised when generalizing beyond the experimental setup. Limitations include a small, geographically concentrated sample, focus on a single ChatGPT model, lack of subtask granularity (e.g., idea generation vs drafting), and EEG’s limited spatial resolution. The authors outline several avenues for future work, including larger diverse samples, multiple LLMs, multi-modal inputs, and fMRI to complement EEG findings.

Quote 3: “As a preprint, peer review is ongoing and results should be treated with caution.” - Nataliya Kosmyna

Future directions

Future research should explore more diverse populations, additional LLMs, and different modalities (audio, visuals). Subtask labeling could clarify how thinking, drafting, and citing unfold under each tool, and integrating fMRI could localize deeper brain structures involved in external-tool use. Longitudinal studies could assess memory retention, creativity, and writing fluency over extended periods.

Conclusion

The article presents a preliminary guide to understanding AI’s cognitive impact on learning environments, highlighting that external AI assistance can reshape cognitive strategies and brain activity in essay writing. The results encourage cautious integration of AI tools in education and call for broader, more rigorous studies to map long-term learning outcomes.