Research shows that over-reliance on AI chat Bots may drop cognitive abilities.

Recently, an in-depth study conducted by the Massachusetts Institute of Technology revealed the potential impact of using large language models (LLMs) such as ChatGPT on cognitive abilities in educational environments. This 4-month study involved 54 participants and aimed to explore the effects of using LLMs on the brain and cognitive abilities while writing articles.

The research team divided the participants into three groups: those using LLMs, those using search engines, and those relying solely on their brains. Participants were required to complete writing tasks on different topics within a limited time. The study employed various methods, including electroencephalogram (EEG) recordings, natural language processing analysis, and participant interviews, to comprehensively assess cognitive engagement and neural activation.

The results show that participants relying solely on their brains exhibited greater diversity in writing styles, while the articles in the LLM group showed a clear trend of homogenization. The LLM group far surpassed the other two groups in using specific named entities (such as names, places, etc.). However, due to time constraints, participants in the LLM and search engine groups often focused more on copying and pasting tool outputs rather than incorporating original ideas.

EEG analysis reveals that with increasing external support, the degree of brain connectivity systematically drops. The group relying solely on brain function exhibits the strongest and most extensive neural network connections, while the overall neural coupling of the LLM-assisted group is the weakest. This finding reflects significant differences in cognitive strategies employed by different groups.

In subsequent experiments, participants who transitioned from LLM to relying solely on their brains exhibited weaker neural connections and lower network engagement. In contrast, participants who shifted from relying solely on their brains to using LLM demonstrated stronger memory recall abilities and reactivated extensive brain regions.

The interview results show that the LLM group has the lowest sense of ownership over the articles they wrote, and they perform poorly in citing content they just wrote. Over 83% of ChatGPT users are unable to accurately cite the content of articles written just a few minutes ago.

Although this study has not yet undergone peer review, its results indicate that during the 4-month research process, participants in the LLM group performed worse in terms of neural, linguistic, and scoring metrics compared to the control group that relied solely on brain function. Researchers point out that the widespread use of LLMs may impact the enhancement of learning skills, especially for younger users.

The research team emphasizes that before confirming the benefits of LLMs for humans, longer-term longitudinal studies are needed to gain a deeper understanding of the long-term effects of AI chat Bots on the human brain. The core findings of this study remind us that while AI tools like ChatGPT may bring convenience, over-reliance may negatively impact cognitive abilities. Therefore, in education and daily use, a careful balance should be maintained between the use of AI tools and the cultivation of independent thinking skills.