June 20 - Time magazine reported on June 18 thatMassachusetts Institute of TechnologyA new Media Lab study finds that over-reliance on ChatGPT likelihoodWeakening of critical thinking skillsThis has led to the emergence of an outside opinion on theGenerative AI Concerns about effects on brain development.
The research team recruited 54 Boston-area residents between the ages of 18 and 39 years old, divided into three groups, each using OpenAI's ChatGPT, Google Search, and Pure HandwritingThe researchers found that participants using ChatGPT were able to complete multiple SAT essays by using EEG. The researchers monitored activity in 32 regions of the subjects' brains via EEG and found that participants who used the ChatGPTLowest brain activity,It is also the worst in terms of neurological, linguistic and behavioral performance.. Over time, the motivation of this group to write declined significantly, and by the late stages most were simply copying and pasting.
The paper notes that the use of large language modelsMay inhibit the learning process, especially affecting young people. Despite the fact that the study is not yet peer-reviewed and the sample size is limited, the paper's author, Natalia Kosmina, has decided to go public early to remind society not to look at immediate convenience at the expense of long-term effects on brain development. "I'm worried that if we wait another six months, policymakers might push for some kind of 'GPT kindergarten' program, which would be extremely harmful, especially to the still-developing brain."
Kosmina has been a full-time research scientist at the lab since 2021. She wanted to focus on the impact of AI in academic settings, as the student body is increasingly using AI. In the study, participants answered SAT writing questions on topics such as the ethics of philanthropy and "choice overload," each with a 20-minute time limit.
The essays of the group using ChatGPT were highly convergent in language and content and lacked independent thinking, as evaluated by two English teachersThese articles are "soulless.". The EEG showed significant deficits in executive control and attentional engagement in this group. By the third essay, many were no longer engaged in the writing, instead leaving the topic directly to ChatGPT. Kosmina recalls, "It's basically 'you give me a piece and I'll just moisturize the sentences'.”
In contrast, the handwriting-only group showsStrongest Brain Connections, especially in the α, θ, and δ bands related to creativity, memory, and semantic processing. Participants showed more engagement and curiosity overall, as well as higher levels of identification and satisfaction with the work.
The Google search group was also active, with brain data and satisfaction at high levels. This is noteworthy, the study noted, as more and more people are choosing toAsk the AI questions directly instead of consulting web pages.
After completing the three writings, everyone is required to rewrite one of them.ChatGPT group this timeThe tool must not be usedChatGPT was available for the first time for the handwriting-only group, while the results showed that the formerAlmost no memory of what you've written, alpha and theta brainwaves were also significantly diminished, indicating that deep memory was barely activated. Kosmina said, "Although the task was accomplished, seemingly quickly and with little effort, we found that theHardly any contentInto their brain's memory system."
On the other hand, in the second group, not only did the quality of writing not deteriorate after AI was enabled, but the neural connections were even significantly enhanced in all frequency bands. This gives us hope:AI can also aid, not hinder, learning if properly guided..
Kosmina calls for more education on AI use, emphasizes 'brain development'Still need analogical training" and advocates for the introduction of more proactive legislative measures to pre-test the use of new technological tools.
Psychiatrist Dr. Zeeshan Khan has been treating children and adolescents for a long time. He notes that many students haveHeavy reliance on AI to complete assignments. "From a mental health perspective, long-term reliance on large language models can have unintended psychological and cognitive effects, with subsequent weakening of neural connections such as information extraction, factual memory, and mental toughness."
Ironically, after the release of the paper, some social media usersSummarized the content in AI and posted it online... Cosmina had expected this to happen and deliberately set traps in the text such asPrompt the AI to "read only the table below" to limit its depth of understanding.
She also found that most of the big language models "make up" a detail out of thin air: she never mentioned the version of ChatGPT she used in her paper, but AI Abstracts, however, claims to be using GPT-4o. She laughed and said, "That's what we were trying to validate, and sure enough the model fantasized on its own."
She revealed that the team is now moving forward with another similar study looking at changes in programmers' brain activity when writing code with or without AI. "So far, the results seem even worse," she said, adding that the study will serve as a warning to companies planning to replace junior developers with AI. Even if efficiency rises, the reliance on AI could lead to a decline in critical thinking, creativity, and problem-solving skills among remaining employees.
Scientific research on the impact of AI is still in its infancy. A May study from Harvard University found that generative AI can improve productivity, and thatBut it will weaken the initiativeThe same month. That same month, MIT had also drawn a line in the sand with a paper written by a student in an economics PhD program that claimed AI could dramatically increase workplace productivity.