Several researchers warn that as large language models take on more cognitive tasks, there will be a cost to pay for this mental externalization.
When researcher Nataliya Kosmyna was looking for interns, she noticed that the cover letters she was receiving were suspiciously similar. They were long, polished, and, after initial introductions, often jumped into making an abstract, arbitrary connection to their work.
It became clear to him that candidates were using large language models (LLM) – a form of artificial intelligence that powers chatbots like ChatGPT, Google Gemini and Claude – to compose their letters.
At the same time, during classes on the Massachusetts Institute of Technology (MIT) campus, Kosmyna – who studies the interaction between humans and computers – began to observe that several students were forgetting content more easily than was the case a few years ago.
Given the growing dependence on LLMs, the professor had the intuition that they could be affecting her students’ cognition and decided to delve deeper into the matter to understand it better.
The concern
The concern for researchers like Kosmyna is that if we come to rely on AI too much, it could affect the language we use and even our ability to perform basic cognitive tasks.
There is now a growing body of research suggesting that this “cognitive download” towards AI can have a corrosive effect on our mental abilities. The consequences could be alarming and even contribute to cognitive decline.
It is well known that the tools we use can change the way we think.
With the advent of the Web, for example, tasks that once required extensive research could be solved by simply entering a simple query into a search box.
As the use of search engines intensified, various research revealed that our propensity to remember details decreased; a phenomenon that has been dubbed “the Google effect.” (Some, however, argue that the Web also acts as an external memory system that frees our brains to engage in other tasks.)
However, there is now growing concern that, as we delegate more and more of our thinking to large language models (LLM) and other forms of artificial intelligence, the effects on our memory and problem-solving abilities may become worse.
AI tools are capable of composing compelling poetry, offering financial advice, and even providing companionship.
Additionally, students are increasingly delegating their own tasks to these AI tools.
Several studies have already shown that young people could be especially vulnerable to the negative effects that the use of AI can have on fundamental cognitive skills, such as critical thinking.
Kosmyna, however, wanted to go even deeper into the analysis of these possible effects.
Reduced mental effort
She and her colleagues at the MIT Media Lab recruited 54 students to write short essays and divided them into three groups.
One was instructed to use ChatGPT. A second group could use the Google search engine, with the AI-generated summaries disabled. The third did not use any technology. Each student’s brain waves were measured while they performed the task.
The essay topics were deliberately stated in an open-ended manner, which meant that the task required very little research; The slogans included questions related to loyalty, happiness or the decisions we make in our daily lives.
The results have not yet been published in a scientific journal, but were nonetheless revealing, according to Kosmyna.
Those who turned only to their own mind showed a brain that was “on fire,” showing widespread activity in many of its areas, the expert said.
The group that only used the search engine showed intense activity in the visual areas of the brain; However, the group that used ChatGPT had noticeably corrupted brain activity: it was reduced by up to 55%.
“The brain did not fall asleep, but there was much less activation in the areas corresponding to creativity and information processing,” says Kosmyna.
ChatGPT also affected participants’ memory. After submitting their essays, the members of the group that used AI were unable to cite fragments of their own texts, and several of them felt that they had no sense of authorship over the work done.
Other studies have also shown that people lose their ability to retain and remember information when using artificial intelligence tools like ChatGPT.
Although the findings are still in the peer review phase, they are similar to those of other studies.
Research by experts at the University of Pennsylvania suggests that some people experience what they call “cognitive performance” when using generative AI chatbots.
This implies that they tend to accept what the AI tells them with minimal scrutiny, and even allow this interpretation to prevail over their own intuition.
Similar effects can be observed outside the realm of AI chatbots, even in life-or-death situations.
A multinational research team recently found that medical professionals who used an AI colon cancer screening tool for three months subsequently showed a reduced ability to detect tumors without the tool’s help.
Delegating work to AI also carries the risk of losing much of the creativity that generates original works, Kosmyna warns.
The essays that the students in her study wrote with ChatGPT turned out to be very similar to each other and were rated by the teachers who evaluated them as “soulless,” lacking originality and depth, Kosmyna notes.
“One of the teachers even asked if the students had been sitting next to each other, given how extremely similar the essays were.”
While studies like this illustrate the short-term effects that large language models (LLMs) can have on the brain, their long-term implications are much less clear.
The study carried out by Kosmyna and her colleagues offers a first glimpse of this.
Four months after the initial study, they asked the students to write another essay; however, on this occasion, those who had used ChatGPT were instructed to work without the support of an LLM.
The neural connectivity in their brains turned out to be corrupted to that of those who had made the transition in the reverse direction, which could suggest that, in the first place, they had not adequately engaged with the topics covered.
Cognitive impairment
Large language models (LLMs) can be a positive tool to stimulate thinking, but only if we do not depend on them by delegating our mental tasks in the process, says computational neuroscientist Vivienne Ming, author of “Robotic Proof“.
However, he worries that this is not the way most people interact with this technology.
His reasoning is based on research he conducted for his book, during which Ming asked a group of UC Berkeley students to predict right-world outcomes, such as the price of oil.
He found that most participants had simply gone to the AI and copied the answer.
He measured gamma wave activity in their brains – an indicator of cognitive effort – and realized that they showed very little activation.
It is worth reiterating that his research has not yet been published; However, Ming is concerned that if his findings are confirmed by subsequent studies, this could have long-term implications.
Other research, for example, has linked weak gamma wave activity to cognitive decline later in life.
“That’s really worrying,” says Ming. “If that becomes the pure way that people interact with these systems – and we’re talking about smart kids – it’s a bad thing.”
Deep thinking, he maintains, is our superpower.
“If we don’t exercise it, the long-term implications for cognitive health are extremely significant.”
This is because very little cognitive effort is required when relying on LLM, but Ming adds that cognitive effort is precisely what a healthy brain needs.
However, a small subset of participants – less than 10% – worked differently and used AI as a tool to collect data that they then analyzed themselves.
These individuals made more accurate predictions than the other participants and also showed greater brain activation.
Almost two decades ago, Ming predicted that within 20 to 30 years, we would be able to see a statistically significant increase in dementia rates, directly related to our over-reliance on Google Maps.
“My intention was to be provocative,” says Ming. “If you don’t have to think about how to navigate, then there will be some detectable effect.”
While we don’t have data on this exact prediction, increased use of GPS has been linked to a decline in spatial memory over time, according to a study of 13 people over three years.
Additionally, poor spatial navigation could be a possible predictor of Alzheimer’s disease, according to another study.
It is evident that the more active our brain remains, the greater its protection against cognitive decline.
Therefore, Ming points out, large language models (LLMs) could not only impair creativity, but also impair cognition and potentially increase the risk of dementia.
As the use of AI tools increases, we must work with them in a way that benefits us rather than harms us.
Ming suggests that the goal could ultimately be a form of “hybrid intelligence” in which humans and machines “tackle difficult tasks” together.
By this, she means that we must think for ourselves first and use the tools later to challenge us, rather than simply allowing them to answer our questions.
Kosmyna agrees with this approach and suggests learning the different subjects without resorting to AI tools in the first stage – in order to lay a solid foundation – and only then consider the use of large language models (LLM).
Ming recommends using what she calls the “nemesis instruction” to test one’s reasoning.
This method involves asking the AI to take on the role of a “bitter enemy” or nemesis, and then asking it to explain in detail why our solutions are wrong and how we could correct them; In this way, we are forced to defend and refine our arguments, instead of limiting ourselves to accepting the answers that the tool offers us.
Another technique she proposes is to prioritize “productive friction,” asking the AI to simply provide context and ask us questions, rather than directly providing us with answers.
When testing this method – by configuring an AI bot to refrain from providing solutions – he observed that users showed a greater degree of involvement and participation.
In short, we should all remain alert to cognitive shortcuts, something that – as Kosmyna points out – “our brain loves.”
Obviously, to ensure long-term brain health, it is common for us to continue to constantly challenge ourselves.
In this process, our mind, our creativity and our cognitive health will benefit.
This is a Spanish adaptation of a story originally published by BBC Tradition. To quiz the English version, do click here.
click here for perhaps more stories from BBC Data Mundo.
Subscribe here to our new newsletter to receive a selection of our best content of the week every Friday.
You can also follow us on YouTube, instagram, TikTok, x, Facebook and in our whatsapp channel.
And remember that you can receive notifications in our app. Download the latest version and activate them.
- “The real danger of artificial intelligence is human stupidity”
- 6 expert tips on how to talk to AI to get better answers
- AI: 4 questions we should ask ourselves before using any artificial intelligence tool