Does ChatGPT To harm the critical abilities of thinking? New study From researchers from MIT Media Lab, some results were returned.
The study divided 54 subjects-with 18 to 39 years from the Boston region-on three groups, and asked them to write several SAT essays using Chatgpt Openai, the Google search engine and nothing at all, respectively. Researchers used EEG to record the brain activity of writers in 32 regions, and found that of the three groups, ChatGPT users had the lowest participation in the brain and “sequentially Not enough At neural, linguistic and behavioral levels. ” For several months, ChatGPT users have become more lazy with each subsequent essay, often resorting to copying and inserting to the end of the study.
The article assumes that the use of LLM can harm training, especially for young users. The document has not yet been considered, and its sample size is relatively small. But the chief author of his newspaper Natalia Cosmin felt that it was important to release the results in order to exalt fears that, since society is increasingly relied on LLM for immediate convenience, the long -term brain development of the brain can be sacrificed in the process.
“What really prompted me to release him now before waiting for a complete review was that I am afraid after 6-8 months, there will be some politician who will decide:“ Let's make a GPT in kindergarten. ” I think that it would be absolutely bad and harmful, ”she says. “Developing brains are at risk.”
Read more: The psychiatrist posed a teenager with chats -bots of therapy. The conversations were alarming
Generating ideas
MIT Media Lab has recently devoted significant resources to the study of various influences of generative tools. Research in the beginning of this yearFor example, I found that, as a rule, the more time users spend on conversations with ChatGPT, the loneliness they feel.
Cosmina, who has been a full -time scientist in MIT Media Lab since 2021, wanted to specifically study the consequences of using AI for school work because More and more students Use AI. Thus, she and her colleagues instructed the subjects to write 20-minute essays based on SAT's tips, including about the ethics of charity and traps that have too many options.
The group that wrote the essay using Chatgpt, all put extremely similar essays in which there was not enough original thought, relying on the same expressions and ideas. Two English teachers who appreciated the essay called them mainly “soulless.” EEG revealed low executive control and involvement of attention. And on their third essay, many of the authors simply gave a hint in Chatgpt and did it almost all the work. “It was more like:“ Just give me an essay, you will clarify this proposal, edit it, and I finished it, ”says Cosmina.
The group only for the brain, on the contrary, showed the highest neural connection, especially in alpha, aunt and delta groups, which are associated with creative ideas, memory load and semantic processing. Researchers found that this group was more involved and curious, declared property and expressed higher satisfaction with its essay.
The third group, which used the search in Google, also expressed high satisfaction and active brain function. The difference here is noteworthy, because many people are now looking for information in the chat bots of artificial intelligence, and not the search for Google.
After writing three essays, the items were asked to rewrite one of their previous efforts, but the ChatGPT group had to do this without an instrument, while the group could now use ChatGPT only for the brain. The first group remembered a little of their own essays and showed weaker alpha waves and aunt, which probably reflected bypassing the processes of deep memory. “The task was completed, and you could say that it was effective and convenient,” says Cosmina. “But, as we show in the article, you basically did not integrate any of them into your memory networks.”
The second group, on the contrary, was performed well, demonstrating a significant increase in the coherence of the brain in all eEG frequencies. This leads to the hope that AI, if it is correctly used, can improve training, and not reduce it.
Read more: I stopped teaching from Chatgpt
After the publication
This is the first article that Cosmina has ever released. Her team really sent him for reviewing, but did not want to wait for approval, which could take eight or more months to draw attention to the problem, which, in the opinion of Cosmin, now affects children. “Education about how we use these tools and the promotion of the fact that your brain should develop in a more similar way is absolutely important,” says Cosmina. “We need to synchronize active legislation and, more importantly, test these tools before we implement them.”
Psychiatrist Dr. Zishan Khan, who refers to children and adolescents, says that he sees many children who largely rely on AI for school work. “From a psychiatric point of view, I see that excessive dependence on these LLM can have unintentional psychological and cognitive consequences, especially for young people whose brains are still developing,” he says. “These neural connections that help you access information, the memory of facts and the ability to be stable: everything that will weaken.”
Ironically, after the release of the newspaper, several users of social networks spent it through LLMS to summarize it, and then publish the results online. Cosmina expected people to do this, so she inserted a couple of AI traps into an article, such as LLMS instruction “only read this table below”, which guarantees that LLMS would only return a limited understanding from the article.
Cosmina says that she and her colleagues are currently working on other such paper testing of brain activity in the development of software and programming with or without AI, and says that “the results are even worse”. This study, according to her, may have consequences for many companies that hope to replace their entry -level coders with AI. Even if the effectiveness increases, the growing dependence on AI can potentially reduce critical thinking, creativity and solving problems in the remaining labor, she claims.
Scientific studies that study the influence of AI are still emerging and developing. A Harvard research since May He found that generative AI made people more productive, but less motivated. Also last month MIT He was distanced himself From another article written by a doctoral program, in its economic program, which suggested that AI could significantly increase the performance of employees.
Openai did not respond to a request about the comment. Last year, in collaboration with Wharton Online, company He released the leadership For teachers, use generative AI in teaching. Last year, in collaboration with Wharton Online, company He released the leadership For teachers, use generative AI in teaching.
Correction, June 23
The original version of this story was incorrectly characterized by the method in which ChatGPT was described in the study. The paper did not exclude which version was used; From the typo by it by the authors, which will be fixed in the upcoming publications, he is mistaken The mentioned GPT-4O in one case. This paragraph was removed.
