AI making people dumb, reveals MIT study

Relying on Large Language Models (LLMs) like ChatGPT to write essays is making people "less smart," according to a new MIT study that tracked participants' neural activity and learning performance over four months. 

The study found that individuals who used only their brains to complete writing tasks consistently outperformed those using AI tools in terms of memory, cognitive engagement, and essay quality.

The research divided participants into three groups: Brain-only (no external help), Search Engine (internet tools like Google), and LLM-assisted (language models like ChatGPT). 

Over the course of four distinct sessions, each group was tasked with writing essays while their brain activity was monitored through EEG (electroencephalogram) analysis. Interviews and memory recall tests followed each writing session.

The study involved over 100 participants, who in the final session were switched – (LLM-to-Brain and Brain-to-LLM) to observe behavioral and neurological changes. 

Researchers found a strong and consistent pattern: Brain-only participants displayed the most robust neural connectivity, especially across alpha and beta wave networks associated with attention and memory. The Search Engine group showed moderate engagement, while the LLM group exhibited the weakest brain network coupling. 

All the same, participants who switched from LLM to Brain mode in the final session showed improved memory and re-engagement of key visual and cognitive processing areas, such as occipito-parietal and prefrontal brain regions.

From a linguistic standpoint, the essays generated by the Brain-only group were more cohesive and conceptually rich, with consistent patterns in Named Entity Recognition, n-grams, and topical structure, the study revealed. 

However, participants relying on LLMs demonstrated significantly lower recall of their written content, and expressed weaker feelings of authorship and ownership during post-session interviews.

"Our findings suggest that over-reliance on LLMs like ChatGPT may erode key learning processes, especially those involving memory, comprehension, and cognitive strategy," the researchers said.

The study did not reveal which version of ChatGPT they used for the experiment.

The paper is yet to be peer-reviewed and the draft published was an attempt on the authors' part to thwart any attempts at including LLMs into early school classes. 

"What really motivated me to put it out now before waiting for a full peer review is that I am afraid in 6-8 months, there will be some policymaker who decides, 'let's do GPT kindergarten.' I think that would be absolutely bad and detrimental," the paper's main author Nataliya Kosmyna said, according to Time. 

"Developing brains are at the highest risk," she said.