Tool or Crutch: ChatGPT’s Effect on Neural Activity and Cognition
By Paige Nicklas
A large-language model walks into a bar.
The bartender asks, “What will you have?”
The LLM says “What’s everyone else having?”
Artificial intelligence and large-language models (LLMs) are now regularly used, with users touting their convenience, creativity, and productivity. This fits into the broader human tendency to take advantage of tools and external methods to reduce effort and boost performance and productivity – a phenomenon known as “cognitive offloading.” Something as simple as keeping a to-do list on a sticky-note can be considered cognitive offloading. This transfers mental tasks from an individual’s mind to external tools or systems to reduce cognitive demand and free up mental resources for other tasks. (Hu et al., 2019). What remains unknown is how unfettered access to the unprecedented “convenience” of LLMs impacts the brain.
The scope of questions regarding the potential impact of offloading cognitive tasks to AI is massive and complex, with vast implications (both positive and negative) for cognition, creativity, education, and psychology. One new study, “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task” (Kosmyna et al., 2025), begins to address these questions by pitting LLM users against non-users on cognitively demanding tasks, like writing an essay, while measuring neural activation. This study has gained heavy media attention, often through strongly worded (clickbait) titles.
In this study from MIT, 54 participants were sorted into three groups and asked to write SAT essays over 3 separate visits across 4 months while the study team used electroencephalography (EEG) to record the participants’ brain activity across 32 regions while they wrote. The three groups were able to write their essays using either OpenAI’s ChatGPT, Google’s search engine, or no external tools at all (“Brain-Only”). 18 participants took part in an optional fourth session where they were reassigned to a group opposite of their initial assignment but wrote an essay on a topic they previously wrote about. The longitudinal nature of the study design allowed the researchers to track changes in the participants’ behavior and brain activity over time.
At the first session, 83% of the participants in the LLM group were unable to provide a quote from the essay they completed just minutes earlier, compared to only about 11% of the participants in each of the other groups. Of those in the LLM group who could provide a quote, none of the quotes were accurate. Quotes were scored based on verbatim recall, not general content accuracy. Across all 3 of the main sessions, these trends continue, with Search Engine and Brain-Only groups demonstrating significantly less issues with quoting ability and accuracy. In session 4, participants that were reassigned from LLM to Brain-only showed a reduction in sustained working memory usage, while participants reassigned from Brain-only to LLM exhibited higher memory recall. Notably, over the course of all the sessions, with each subsequent essay, the LLM group was more likely to resort to simple copy-and-paste by the end of the study.
To study brain activity, the EEG recordings in each group were compared based on a measure of neural connectivity. The study team analyzed how coordinated the activity between the 32 recording channels was within each of the 3 groups, across the 4 commonly studied frequency bands: alpha, beta, theta and delta. These bands in EEG refer to different frequencies of the waveform of the recorded electrical brain activity, and each band is associated with different cognitive and attentional states and processes (Nayak & Anilkumar, 2023). The current study provides general overviews of functions associated with each band, though it should be noted that there is variation in the broader literature. The authors note that alpha bands and beta bands are typically associated with internally driven planning, critical reasoning, and working memory, while theta oscillations are thought to be related to encoding and retrieval, needed for the integration of information. Delta bands are less studied, but associated with semantic retrieval, creative ideation, and coordination of complex cognitive processes across large networks.
Across all frequency bands, the Brain-only group had consistently greater connectivity between brain regions. The authors interpret these findings by suggesting there’s a difference between entirely internal (“Brain-only”) processing and production of information, compared to a more hybrid approach like in the Search Engine or LLM groups. With access to external information and processing, particularly in the LLM group, participants are able to offload cognitive engagement and effort. This isn’t without consequence, though, as evidenced by their inferior memory of what they actually wrote. It’s possible that lesser connectivity interrupted the typical memory encoding and formation processes, and metacognition that occur when one is fully and critically engaged in a cognitive task. Further, more complex analyses in this study also suggested that this lack of engagement led to the LLM group having a narrower range of ideas incorporated into their essays, less creativity, and a weaker sense of ownership over the final product.
This paper is vast, and still only a preprint at the time of writing this commentary. Once it completes the peer-review process, its conclusions can hold even more substantial weight. (I do exalt the reviewers who will take on this 206-page, 92 figure manuscript). Some notable limitations include sample size, especially at the fourth visit. Further, participants only had 20 minutes to craft responses which may have added time pressure leading them to copy/paste directly from LLM responses. Currently, the results confirm what we’ve known for a long time: the brain is like a muscle, and many cognitive, social, and emotional skills often operate under a “use it or lose it” approach. Relying on LLMs may be pushing us toward the ‘lose it’ end with unprecedented access and speed.
One (perhaps agitational) opinion from a commentary published on Medium posits “Will the use of LLM force us to visit a new kind of gym for the brain to keep it in shape, as, out of our laziness, we will use it less and less?” (Matile, 2025). While this may be catastrophizing the situation a bit, results from this study and similar ones should not be ignored. Though, others argue that there is value in the enhancement AI can bring to human users, with one writer stating that if you compared people doing long division as the study authors did with people writing essays, “…the same would happen if students were asked to perform complex calculations with or without a calculator. The group doing calculations by hand would sweat, while those with calculators would barely blink an eye.” (Kovanovic & Marrone, 2025)
We cannot ignore the complexing of impacts surrounding LLM use.
We cannot ignore the complexing of impacts surrounding LLM use. Thus, perhaps the most important take-away at this moment is the big remaining question: What are we going to do about it?
In the US, there are currently 31 proposed bills at the federal level surrounding AI use, most of which are classified as “Mitigating Harms” legislation (American Action Forum, June 2025). A wide range of proposed, enacted and failed legislation at the state level (for a great tracker of every state, see this webpage and this summary). The European Union, Canada, South Korea, and others have also begun implementing their own forms of AI regulation and policy, all taking different approaches to try and balance the evident utility of these tools with their pervasive negative implications if exploited. The study also importantly acknowledges the vastly higher energy expenditure of LLM uses compared to the Search Engine group, and estimated their expenditures based on a 2023 study (de Vries) – though this is an entirely additional issue. As usual, it is difficult to align policy and research timelines as both need to be reciprocally proactive and reactive.
This study begins to scratch the surface of the work that needs to be done to comprehensively understand the impacts of AI and LLMs on human behavior, cognition, and interaction. Although it has not been published in its final form yet, releasing it in the pre-print version allows it to be a living document and spark conversations on the important findings and to proactively inform policy and education decisions. It is already doing so in mainstream news outlets, like this piece from Time.
We are currently at a crossroads, and it is imperative that we conduct and interpret this and similar research very carefully so we can best inform future decisions across all domains of everyday functioning. The line between enhancement and exploitation is growing ever thinner and blurrier.
What would you like to read about next? Let us know here!
~~~
Written and illustrated by Paige Nicklas
Edited by Faustina Jayaprakash and Tony Nguyen
~~~

References
de Vries, A. (2023). The growing energy footprint of artificial intelligence. Joule, 7(10), 2191-2194. https://doi.org/10.1016/j.joule.2023.09.004
Hu, X., Luo, L., & Fleming, S. M. (2019). A role for metamemory in cognitive offloading. Cognition, 193, 104012. https://doi.org/10.1016/j.cognition.2019.104012
Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X. H., Beresnitzky, A. V., … & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv preprint arXiv:2506.08872
Kovanovic, V. & Marrone, R. (2025). MIT researchers say using ChatGPT can rot your brain. The truth is a little more complicated. The Conversation. https://theconversation.com/mit-researchers-say-using-chatgpt-can-rot-your-brain-the-truth-is-a-little-more-complicated-259450
Matile, P. (2025). Will our brain capability survive LLM development? Medium. https://open-borders.medium.com/will-our-brain-capability-survive-llm-development-1bf855ad220e
Nayak, C. S., & Anilkumar, A. C. (2023). EEG Normal Waveforms. In StatPearls. StatPearls Publishing.


