According to research on new Prilints from MIT, using AI chatbots can reduce activity in the brain, rather than achieving the same task as accomplishing the same task.
Aiming to try to understand how the use of LLM chatbots affects the brain, a team led by research scientist Dr. Natalya Kosmina at MIT Media Labs connected a group of Boston-area university students to an EEG (EEG) headset and gave them 20 minutes to write a short essay. One group was instructed to write without external assistance, the second group was allowed to use search engines, and the third group was instructed to write with the support of Openai's GPT-4O model. This process was repeated four times over a few months.
Although no peer reviews have been reviewed yet, pre-published research findings suggest a significant difference between brain activity in the three groups and the creation of corresponding neural connectivity patterns.
Frankly, visually speaking, brain activity in the LLM-used cohort was… a bit dim.
Looking at brain activity in three study cohorts (left to right: LLM, search, brain group) turns red in colour and activates DDTF size – Click to enlarge
EEG analysis showed that each group exhibited a different pattern of neural connectivity, indicating that brain connectivity was “systematically reduced by the amount of external support.” In other words, search engine users were less involved in the brain, and the LLM cohort “induced” as “the weakest overall coupling.”
Participants' cognitive load was measured using a method known as the dynamically oriented transfer function (DDTF), which measures specific brain activity associated with information flow across different brain regions. According to researchers at MIT, DDTF can explain the strength and direction of flow, making it an appropriate expression of “executive function, semantic processing, and attentional regulation.”
The researchers said the search engine group showed DDTF connectivity of 34-48% compared to the baseline established by the group writing using only gray and white. Meanwhile, the LLM group showed that the magnitude of the DDTF signal was reduced by up to 55%.
Simply put, relying on LLMS, and not so much, search engines – significantly reduce task-related brain connections and show a decline in cognitive engagement during essay writing tasks.
“The brain-only group utilized widely distributed neural networks for internally generated content,” the researchers said of the results. “Search engine groups relied on hybrid strategies of visual information management and regulatory control, and LLM groups optimized for procedural integration of AI-generated proposals.”
As the researchers have explained, these distinctions have “important implications” in how we understand educational practices and learning, that is, there appears to be a clear trade-off between internal integration of information and external support.
Participants in the LLM group were worse than counterparts in the brain-only group at all levels
The LLM cohort discovered by the research team, which tested perceived ownership as recalls of what they wrote was clearly worse.
“This study illustrates the pressing issue of a higher likelihood of lowering learning skills,” the researchers said. “Participants in the LLM group were worse than counterparts in the brain-only group at all levels.”
The fourth session of essay writing enhanced these findings. In the previous research session, participants who were originally told to rely on the brain or LLMS exchanged roles and were given instructions for another essay. Naturally, LLM groups performed poorly when asked to rely on their own thought processes.
“In Session 4, AI support was removed, causing participants to be significantly damaged from the original LLM group,” the researchers said. The opposite was true for other cohorts. “The so-called brain-to-LLM group showed a significant increase in brain connectivity across all EEG frequency bands when allowed to use LLM on familiar topics.”
AI Chat: Do you generally eliminate thoughts?
Findings suggest that AI use early in the learning process could “result in shallow encoding.” This can lead to insufficient recalls of facts and lack of learning, as all the efforts are offloaded. Learn about something using the Faculty of Cognition, and use AI to further research skills to become fully accepted.
“By summarizing, these findings support an education model that slows AI integration until learners engage in adequate self-directed cognitive efforts,” the MIT team concluded. “This approach could promote both the effectiveness of immediate tools and sustained cognitive autonomy.”
That may not be a shocking conclusion, but given the growing number of young people relying on AI to do their studies, the problem needs to be addressed before the world produces an entire generation of intellectually developed AI junkies.
Cosmina said Register She claims she doesn't want to use words like “silly, stupid, or brain rot” to refer to us the effects of AI, claiming it would harm the work her team has done. Still – it has an effect that needs to be addressed.
“These tools provide unprecedented opportunities to enhance learning and access to information, but their potential impact on cognitive development, critical thinking and intellectual independence require very careful consideration and ongoing research,” the paper concluded.
As the paper has not yet undergone peer review, Cosmina noted that the conclusion should be “treated carefully and preliminary.” Nevertheless, she writes that the findings before the review remain useful “as a preliminary guide to understanding the cognitive and practical impacts of AI on learning.”
The team hopes that future research will consider the impact of AI on memory retention, creativity and written flow ency, as well as the use of LLM in modalities beyond textbooks.
Regarding the next plan to investigate by the MIT team, Kosmyna said the team is paying attention to considering coding atmosphere and using AI to generate code from natural language prompts.
“We're already collecting data and are currently working on analysis and drafting,” Kosmyna says. She added that the work “expects additional research to be triggered in a variety of protocols, populations, tasks and methodologies, as this is one of the first such studies to study the effects of AI on the human brain.
AI will creep into every apparent aspect of our lives at an ever-growing pace, so there will be much work to do. ®
