Created:2025-10-05 10:01 Tags: ai cognition brain_scan neuroscience
🖼️ Image to restore:
Gemini_Generated_Image_n6qwwzn6qwwzn6qw.png
(Image: created by Gemini)
This is a rather belated reflection on a hot paper that came out in early July from an MIT research team (Kosmyna, et al., 2025). They preprinted a paper exploring the impact of AI on human brain activity by scanning brains while people wrote with ChatGPT assistance versus writing purely with their own brains. The draft of this review has been sitting around for a while, only making its way out in early October. Part of the delay was my ongoing job search, but also my slow journey into the “sharing is writing” philosophy of a Digital Garden. Letting go of the “heavy” notions of formal writing and publishing isn’t easy. I might share more about that experience on other channels sometime.
The researchers randomly assigned 45 participants into three groups: one using ChatGPT for writing, another using traditional search engines, and a third relying solely on their own brains. Each group completed three 20-minute writing sessions, choosing one topic from three options per session. Finally, 18 of the 45 participants volunteered for a fourth session, where they wrote using a method opposite to their first three sessions (e.g., an LLM-assisted writer would do brain-only writing).
The study’s conclusion (summarized by GoogleNotebook and reviewed by me) was: “While LLMs can make tasks easier, excessive reliance on them may hinder the development of crucial cognitive skills, such as critical thinking, deep memory encoding, and a sense of personal ownership over intellectual work.” This immediately brought to mind two related papers: Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance (Fan et al., 2025) and Wang and Fan’s (2025) The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: Insights from a meta-analysis. I’ll dive into those two and discuss them together when I’ve had a chance to read them thoroughly.
This conclusion has been widely cited on LinkedIn as further “proof” of the negative cognitive and neural impacts of relying on ChatGPT. However, several individuals have also offered in-depth critiques of the study. For instance, these three pieces:
The latter two, in particular, are quite strongly worded, pointing out serious issues with the paper, especially regarding the small sample size, methodological flaws, and overly assertive conclusions.
My understanding of brain science is, to put it mildly, rudimentary. So, I can’t really evaluate or comment on the methodology, theory, or terminology of the paper. However, when I read the following two paragraphs, a big question mark popped up in my mind regarding the study design:
“Across the essays written by different groups one can likely observe propagation of biases used in the training data of the used LLM, or advertisements in a search engine, or human biases,” (p. 133) The authors found that writers using AI or search engines were influenced by biases in training data and advertisements.
Another paragraph states, “Participants in the LLM and Search Engine groups were more inclined to focus on the output of the tools they were using because of the added pressure of limited time (20 minutes). Most of them focused on reusing the tools’ output, therefore staying focused on copying and pasting content, rather than incorporating their own original thoughts and editing those with their own perspectives and their own experiences.” (p. 134) The authors observed that within the 20-minute writing period, writers using AI and search engines tended to output content they found, showing less originality compared to those writing purely with their brains.
At first glance, it’s easy to think, “Oh, so human thought is indeed influenced by LLMs and search engines.” But reflecting on my own writing experience, a question arose: Why would the “brain-only” writers be different? Could this issue stem from the design of the 20-minute writing task itself?
In my writing experience, the most challenging phrase is trying to integrate what I’ve read into my own arguments, especially in literature reviews. It’s so easy to get swept away by the already-constructed viewpoints and structures of the literature, making my own writing fragmented. Whenever I find myself in this predicament, I need to step away from the source materials and rethink what I truly want to say and how to structure it, often in a blank document.
In the Kosmyna team’s study, the “brain-only” group had no access to external information, unlike the AI tool and Google groups. Therefore, it’s reasonable to assume their brain activity would naturally be different. If they were given a stack of physical reference materials, wouldn’t they also be influenced by those?
From this perspective, I see a problem with the study’s design: AI and search engines, as tools, do more than just modify text; they link to a massive database outside the human brain. Their knowledge base is undoubtedly far greater than what a human brain can memorize. So, for the LLM and search engine groups, the 20 minutes actually included steps for gathering and decoding information, then combining that information with personal knowledge and experience to form writing material. For the “brain-only” group, these two steps were completed before the experiment began; participants were drawing from their past experiences and knowledge. From a sociological standpoint, all of a participant’s past experiences and knowledge are also influenced by various forms, and cannot simply be labeled as “original” or “creative.”
Therefore, for stricter research design, the “brain-only” group shouldn’t have been just brain-only. Instead, during the 20 minutes of writing, the researchers should have provided them with materials through a closed medium (like paper) and scanned their brain states while they collected, analyzed, and wrote with that data. This would make the data from all three groups truly comparable.
Scanning the brain is a great idea. In the months since this paper was published, its conclusions have been widely cited on social media, even extending into the Chinese internet world, including by some content creators I admire. However, almost everyone has paired the study’s conclusion with the “MIT” label as yet another piece of “evidence” for “AI anxiety.” This has made me incredibly curious about the current state of knowledge mobilization in academic research and ponder what role academic research can truly play in building social consensus in the age of self-media.
References： Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
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.
Wang, J., & Fan, W. (2025). The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: Insights from a meta-analysis. Humanities and Social Sciences Communications, 12(1), 621. https://doi.org/10.1057/s41599-025-04787-y