Brain’s performance during, after exclusive ChatGPT usage drops, results in weaker neural connectivity

While exclusively relying on ChatGPT showed concerning outcomes, participants who wrote essays without any assistance and then began using ChatGPT showed higher memory recall, efficient tool use, strong re-engagement and enhanced neural connectivity. In contrast, participants who first used ChatGPT to write essays and then moved to unaided writing underperformed cognitively

Preliminary findings from a small study involving 54 participants from five universities from the greater Boston area who were equally split into three groups — ChatGPT group, Search Engine group, and Brain-only group — found that participants who exclusively relied on ChatGPT large language model (LLM) to write essays had the lowest brain engagement as reflected by the weakest neural connectivity — up to 55% lower in low-frequency networks — and also displayed diminished critical thinking skills and memory recall.

On the other hand, the brain-only group participants, who were not permitted to use any LLM and online websites, displayed highest neural connectivity, especially in alpha, theta, and delta bands, which are the core zones for focus and memory. They also displayed better cognitive engagement, stronger executive control, and deeper creative processing. Participants in the Search Engine group displayed better neural connectivity than the ChatGPT group, and had better and comparable critical thinking skills and memory recall as the group that did not rely on any external assistance to write the essays.

The participants were required to attend three writing sessions and be interviewed by the authors after each session, while the fourth session was optional. In the fourth session, participants switched groups: participants in the LLM group wrote an essay without using any tool, while the Brian-only group used ChatGPT.

The study was carried out by MIT researchers over four months. The results have been posted as a preprint, which are yet to be peer-reviewed.

Timing matters

While exclusively relying on ChatGPT showed concerning outcomes, participants who wrote essays without any assistance during the first three sessions and then began using ChatGPT in the fourth session showed higher memory recall, efficient tool use, strong re-engagement and enhanced neural connectivity, which was much higher than when they were writing without using ChatGPT. In contrast, participants who moved from ChatGPT to unaided writing underperformed cognitively with reduced alpha, beta activity and poor content recall; they had potential limitations in achieving robust neural synchronisation essential for complex cognitive tasks.

However, the fourth session had only 18 participants who were split into two groups with just nine participants per group. A larger study involving a higher number of participants from diverse backgrounds is necessary, particularly for the fourth session, to confirm the findings of this study.

Cognitive debt

The study found that essays written by participants in the ChatGPT group were “biased and superficial” with low memory recall as they failed to critically engage with the essay topic. This pattern leads to accumulation of cognitive debt — a condition in which repeated reliance on external systems like LLMs replaces the effortful cognitive processes required for independent thinking. “Cognitive debt defers mental effort in the short term but results in long-term costs, such as diminished critical inquiry, increased vulnerability to manipulation, decreased creativity. When participants reproduce suggestions without evaluating their accuracy or relevance, they not only forfeit ownership of the ideas but also risk internalising shallow or biased perspectives,” they write.

Besides studying neural connectivity using EEG, the authors during the interviews following each session assessed how participants’ memory, quote recall and ownership of the essays. To evaluate participants’ memory, they were asked to quote from memory any sentence from the essay and to summarise the main points or arguments made in the essay.

Superior performance

Quoting ability was highest among the brain-only group, closely followed by the Search Engine group. LLM users significantly underperformed in quoting ability — 83% (15/18) had difficulty quoting in the first session with none providing correct quotes. It improved in the second session; but even in the third session, six out of 18 participants failed to quote correctly.

Even in the case of being able to quote correctly, the LLM group fared the worst — not one participant could quote correctly after the first session; there were “persistent impairments” in the second and third sessions too. The authors say that this suggests that memory encoding was shallow and the semantic content was not fully internalised. The Brain-only group claimed full ownership of their work, closely followed by the Search Engine group. However, LLM group participants expressed either no ownership or partial ownership, which suggest a diminished sense of cognitive agency. Brain-only participants focussed more about “what” and “why” they wrote; LLM users focused on “how”.

If Brain-only group participants displayed “strong variability” in the way they approached essay writing, the content of the essays written by LLM group participants was “homogeneous” with significantly less deviation compared with other two groups. The Search Engine group was influenced by the content that is optimised by a search engine; the keywords used were likely influenced by the participants’ own queries.

Neural connectivity

Participants in the three groups had significantly different neural connectivity patterns, reflecting divergent cognitive strategies. Brain connectivity of participants was influenced by the amount of external assistance. While the Brain‑only group showed the “strongest, widest‑ranging networks”, Search Engine group displayed “intermediate engagement”, and LLM assistance elicited the “weakest overall coupling”. Brain-only group showed the highest neural connectivity, especially in alpha, theta, and delta bands, followed by the Search Engine group, which demonstrated 34-48% lower total connectivity in lower frequencies. LLM users had the weakest connectivity, up to 55% lower in low-frequency networks. Despite being under greater cognitive load, the Brain-only group demonstrated “deeper learning outcomes and stronger identity with their output”. The LLM group showed “weaker memory traces, reduced self-monitoring, and fragmented authorship”.

Unintentional consequences

This trade-off highlights an important educational concern: AI tools may “unintentionally hinder deep cognitive processing, retention, and authentic engagement with written material. If users rely heavily on AI tools, they may achieve superficial fluency but fail to internalise the knowledge or feel a sense of ownership over it”.

“These findings support an educational model that delays AI integration until learners have engaged in sufficient self-driven cognitive effort. Such an approach may promote both immediate tool efficacy and lasting cognitive autonomy,” the authors write. The authors say that based on low neural connectivity and behavioural quoting failure in LLM group’s participants, there is evidence that early AI reliance may result in shallow memory encoding. This can be prevented if LLM tools during early stages are not allowed, which might support memory formation.