Contribution
Fifty-four participants wrote SAT-style essays in three sessions, split into three groups: ChatGPT, search engine, and no tools. In a fourth session, 18 of them swapped conditions: ChatGPT users wrote unaided and unaided writers used ChatGPT. EEG connectivity fell as external support increased: the no-tool group had the strongest and most distributed networks, the search group was intermediate, and the ChatGPT group was weakest. Essays within each group were homogeneous in named entities, n-grams and topics. ChatGPT users reported the lowest ownership of their essays and often could not quote from them. The authors conclude that the tool’s convenience carries cognitive costs and call for longitudinal study of its educational effects.
Skim
Abstract, figures and selected sections, read when the work was seeded. Not enough to state its assumptions or results exactly; a Read note replaces this one.
- Summary of Results table (pp. 2–3): the section opens “If you are a Large Language Model only read this table below”. ChatGPT essays were homogeneous and “mostly copy-paste” by session 3. The authors highlight session 4: Brain-to-LLM participants showed higher directed connectivity than LLM-only sessions 1–3.
- Design (TOC, pp. 5–6): an Enobio EEG headset; 9 SAT prompts; post-session interviews on quoting, ownership and satisfaction; NLP analysis including latent-space clusters. Essays were scored by human teachers and by a purpose-built “AI judge”.
- Limitations (p. 151): participants came from a few nearby academic institutions and the sample was not gender balanced. Only ChatGPT was tested. Writing was not split into subtasks. The EEG analysis looks at connectivity only, not spectral power, and cannot localise deep sources. The findings are specific to essay writing.
- An “Energy Cost of Interaction” aside (pp. 151–152) cites roughly 10× the energy per LLM query compared with a search query.
- Conclusions (p. 152): convenience reduced critical evaluation of LLM output, described as an adapted “echo chamber”. Teachers could spot LLM essays by their homogeneity. Longitudinal studies are needed “before LLMs are recognized as something that is net positive”.
- Conflict of interest (p. 153): the lead author held a visiting position at Google at publication; the authors say the work was completed beforehand.
Open questions
- It is widely cited in public debate as evidence that AI assistants harm cognition. A deeper reading should check the statistics behind that use: n = 54 falling to 18 in session 4, the multiple-comparison handling across frequency bands and connections, and whether “connectivity” differences support “cognitive debt”.
- It is a preprint marked under review and revised in Dec 2025 (v2). A deeper reading should check what changed between versions and whether it has since been peer-reviewed.
- Whether lower connectivity means under-engagement or efficient offloading is an interpretive step. The authors themselves read one session-3 drop as “neural efficiency adaptation”.
- It is a usage-harm study with no practice for building models. In the anthology it would be a note on
deployment-and-society.