New Delhi: Using large language models (LLM) such as ChatGPT for writing essays can have a detrimental effect on brain activity, scientists have found.
A brain-mapping study conducted by scientists at Massachusetts Institute of Technology (MIT) in the US showed that people who used ChatGPT for an essay-writing task had less brain connectivity, used their alpha and beta networks less, and had lower memory when compared to those who did not.
The study paper is pre-print and has not yet been peer-reviewed. It was published by MIT’s Media Lab on 10 June. A pre-print paper is an unedited version of an academic paper that is published prior to peer review and journal publishing.
The group of researchers conducted the essay-writing experiment with 54 participants, dividing them into three groups: those who could use AI tools like ChatGPT, those who could use only search engines, and those who could only use their brains while writing essays. They did this over three different sessions with the same assignments for the three groups, to understand which tool helped produce what kind of essays.
In the fourth session, the researchers asked the group using AI to write an essay using just their brain, and asked the group using just their brains so far to now use AI, to test how their brains had adapted over the last few sessions.
Throughout the sessions, they conducted electroencephalogram (EEG) tests of the participants.
The researchers found that those using AI had significantly lowered cognitive performance as compared to those who were using just their brains for the task.
“AI tools, while valuable for supporting performance, may unintentionally hinder deep cognitive processing, retention, and authentic engagement with written material,” reads the paper.
While the researchers have submitted their paper for peer review, according to Time magazine, they decided to publish the paper earlier seeing it as an “issue that is affecting children now”. Peer review for academic papers usually takes between 6 and 9 months.
Additionally, while the experiment itself used ChatGPT, the paper clarified that the study looked at the broader gamut of LLMs that provide generative content for users.
The study was also only conducted in a small sample group of people, aged 18 to 39, in the Boston area. The researchers admit that their sample size was quite small and future studies would need to involve a wider group of participants, but given the fast adoption of AI in learning and teaching, the need for studies like this is important.
Also Read: Indians biggest consumers of AI-generated news & most comfortable with it—Reuters Institute report
Experiment & findings
According to the study findings, whenever a topic, such as philanthropy, art, loyalty, was given to the participants, the kind of essays produced by those using “brains-only” showed a lot of heterogeneity, leading to a variety of different essays. The LLM-group, on the other hand, produced very similar essays using similar keywords too.
The EEG tests helped show the differences between how each participant’s brain worked with external support like search engines and LLMs, and without.
One of the interesting things the paper noted was that for people using search engines, the parts of their brain that react to visual cues and process that information was quite active. This is because they were using a screen to get information, and had to make sense of it quite often.
But the same parts were not activated in people using LLMs, because they didn’t even need to interact with the screen for too long. LLMs like ChatGPT give information in a ready-made manner, which reduces brain activity even more.
The paper emphasises that using ChatGPT doesn’t just impact how people organise tasks in their daily life, but also impacts the overall brain functioning that does those tasks.
Beyond that, the paper attempted to see if these underlying changes in brain activity were also visible in the behaviour of participants. They conducted interviews with participants about their essays and found that those using just their brains could quote from and recall their essays much better than those who used ChatGPT.
“The most consistent and significant behavioural divergence between the groups was observed in the ability to quote one’s own essay,” said the paper. “LLM users significantly underperformed in this domain, with 83% of participants reporting difficulty quoting their own paper.”
(Edited by Nida Fatima Siddiqui)