Getting to Know the Study
Two years after the Massachusetts Institute of Technology (MIT) reported and touted the great potential of generative AI in the corporate world, the institute’s researchers are now back with another study that reveals a complex picture of ChatGPT’s impact on human cognition. According to the research, there was a pattern of cognitive decline amongst ChatGPT/Generative AI users who were in the study.
ChatGPT, developed by OpenAI, has quickly become one of the most widely used AI assistants in the world. People use it to write emails, summarize documents, generate ideas, and even help with school assignments. But as its use grows, people are asking important questions: Does ChatGPT make us smarter and more efficient? Or does it make us too dependent on technology?
MIT’s latest research provides some answers and accentuates concerns that have been raised for a while now.
How They Ran the Experiments
In a landmark study titled “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,” Dr. Nataliya Kosmyna from MIT’s Media Lab examined the neurological impact of ChatGPT usage on 54 participants aged 18-39 who were all picked from universities located in Boston.
The study, which lasted for about four months, divided participants into three groups: ChatGPT users (LLM group), Google search users, and users who wrote without any tools (brain-only group). Using electroencephalography (EEG) to monitor brain activity across 32 regions, researchers tracked participants as they wrote SAT-prompted essays over multiple sessions, based on the ethics of philanthropy and the difficulty of having many and conflicting choices.
In turn, essays from every user were analysed with the help of Natural Language Processing (NLP), which is a subfield of artificial intelligence (AI) that focuses on enabling computers to understand, interpret, and generate human language. Teachers and an AI Judge were also involved in this process.
Over the course of several sessions, they switched tools between the participants to see how their brains responded to different kinds of support. According to the study, the brain-only group showed the strongest and most active brain connections. Those using Google search had moderate brain activity, while the ChatGPT group had the weakest brain engagement overall, where they “consistently underperformed at neural, linguistic, and behavioral levels,” the study said.
They also found that ChatGPT users became less involved in the writing process, as many simply resorted to copying and pasting the AI’s suggestions, showing less effort and creativity.
More importantly, in the fourth part of the study, there was a reshuffling that involved reassigning LLM users to independently write their essays. As such, LLM users were reassigned to the Brain-only group (LLM-to-Brain), and the Brain-only users were reassigned to use LLM (Brain-to-LLM) to complete the assignment.
In this reassignment, where they were all asked to re-write one of their essays, LLM-to-Brain users could hardly re-write an essay they had previously written with the help of an AI-powered chatbot. “LLM-to-Brain participants showed reduced alpha and beta connectivity, indicating under-engagement,” the paper said.
On the other side, Brain-to-LLM users had a higher memory recall, and produced an even better re-written essay. According to the paper, they showed “higher memory recall and activation of occipito-parietal and prefrontal areas,” which are associated with working memory, pattern recognition, problem solving, and critical thinking. The associated brain regions are crucial for holding information in mind, manipulating it, and accessing past experiences. The Search Engine users also had similar results, where they actively remembered their former essays and could easily claim ownership.
Why These Findings Matter
This test, especially the group reassignment, suggests that the usage of AI-powered chatbots could significantly hinder the development of skills such as critical and independent thinking, memory overload, as well as long-term brain development.
According to Time, the study was not subjected to a peer review before its release. Being Kosmyna’s first pre-review paper, she admitted that she didn’t want to take the chance of waiting for a seven or eight-month period before raising attention to an issue she believes is affecting children drastically.
“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,” Kosmyna tells Time. “Developing brains are at the highest risk.”
What’s Next for AI and Our Minds
Upon the paper’s release, something ironic occurred. It was discovered that several social media users had run the paper through LLMs for summarised insights to post on the internet. In the interview with Time, Kosmyna says it was not surprising as she specifically set traps into the paper hoping that people would fall for it. One of the traps was subtly instructing LLMs to “only read this table below,” in the paper, ergo ensuring that LLMs would return only limited insight from the entire study.
However, this doesn’t entirely mean people were lazy to read through the entire paper, although it might suggest that everyday people are outsourcing their reading capabilities to a technology and downright taking its suggestions or answers without critically thinking through it — exactly what Kosmyna tries to establish in the study.
The study also complements the recent AI dependence debate that occurred after ChatGPT suffered a 12-hour outage that resulted in the disruption of business workflows across the globe. As such, the questions still stand — Are we going to embrace some sort of balance using this new technology while retaining our own unique capabilities? Or are we going to continue outsourcing any and every thing to AI-powered chatbots?
