Koen Van Belle, a test automation engineer who codes for a living, had been using the artificial intelligence large language model Copilot for about six months when one day the internet went down. Forced to return to his traditional means of work using his memory and what he had decades of experience doing, he struggled to remember some of the syntax he coded with.
“I couldn’t remember how it works,” Van Belle, who manages a computer programming business in Belgium, told Salon in a video call. “I became way too reliant on AI … so I had to turn it off and re-learn some skills.”
As a manager in his company, Van Belle oversees the work of a handful of interns each year. Because their company has limits on the use of AI, the interns had to curb their use as well, he said. But afterward, the amount and quality of their coding was drastically reduced, Van Belle said.
“They are able to explain to ChatGPT what they want, it generates something and they hope it works,” Van Belle said. “When they get into the real world and have to build a new project, they will fail.”
Since AI models like Copilot and ChatGPT came online in 2022, they have exploded in popularity, with one survey conducted in January estimating that more than half of Americans have used Copilot, ChatGPT, Gemini or Claude. Research examining how these programs affect users is limited because they are so new, but some early studies suggest they are already impacting our brains.
“In some sense, these models are like brain control interfaces or implants — they're that powerful,” said Kanaka Rajan, a computational neuroscientist and founding faculty member at the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University. “In some sense, they're changing the input streams to the networks that live in our brains.”
In a February study conducted by researchers from Microsoft and Carnegie Mellon University, groups of people working with data worked more efficiently with the use of generative AI tools like ChatGPT — but used less critical thinking than a comparator group of workers who didn't use these tools. In fact, the more that workers reported trusting AI’s ability to perform tasks for them, the more their critical thinking was reduced.
Another 2024 study published last year reported that the reduction in critical thinking stemmed from relying on AI to perform a greater proportion of the brain work necessary to perform tasks in a process called cognitive offloading.
"In some sense, these models are like brain control interfaces or implants — they're that powerful."
Cognitive offloading is something we do everyday when we write our shopping list, make an event on the calendar or use a calculator. To reduce our brain’s workload, we can “offload” some of its tasks to technology, which can help us perform more complex tasks. However, it has also been linked in other research to things like having a worse memory.
As a review published in March concluded: “Although laboratory studies have demonstrated that cognitive offloading has benefits for task performance, it is not without costs.” It’s handy, for example, to be able to rely on your brain to remember the grocery list in case it gets lost. So how much cognitive offloading is good for us — and how is AI accelerating those costs?
This concept is not new: The Greek philosopher Socrates was afraid that the invention of writing would make humans dumber because we wouldn’t exercise our memory as much. He famously never wrote anything down, though his student, Plato, did. Some argue Socrates was right and the trend is escalating: with each major technological advancement, we increasingly rely on tools outside of ourselves to perform tasks we once accomplished in-house. Many people may not perform routine calculations in their head anymore due to the invention of the calculator, and most people use a GPS instead of pulling out a physical map or going off physical markers to guide them to their destination.
Want more health and science stories in your inbox? Subscribe to Salon's weekly newsletter Lab Notes.
There is no doubt these inventions have made us more efficient, but the concern lies in what happens when we stop flexing the parts of the brain that are responsible for these tasks. And over time, some argue we might lose those abilities. There is an old ethos of "use it or lose it" that may apply to cognitive tasks as well.
Despite concerns that calculators would destroy our ability to do math, research has generally shown that there is little difference in performance when calculators are used and when they are not. Some have even been critical that the school system still generally spends so much time teaching students foundational techniques like learning the multiplication tables when they can now solve those sorts of problems at the touch of a button, said Matthew Fisher, a researcher at Southern Methodist University.
On the other hand, others argue that this part of the curriculum is important because it provides the foundational mathematical building blocks from which students learn other parts of math and science, he explained. As Fisher told Salon in a phone interview: "If we just totally get rid of that mathematical foundation, our intuition for later mathematical study, as well as just for living in the world and understanding basic relationships, is going to be off.”
Other studies suggest relying on newer forms of technology does influence our brain activity. Research, for example, has found that students’ brains were more active when they handwrote information rather than typing it on a keyboard and when using a pen and paper versus a stylus and a tablet.
Research also shows that “use it or lose it” is somewhat true in the context of the skills we learn. New neurons are produced in the hippocampus, the part of the brain responsible for learning. However, most of these new cells will die off unless the brain puts effort and focus into learning over a period of time. People can certainly learn from artificial intelligence, but the danger lies in forgoing the learning process to simply regurgitate information that it feeds us.
In 2008, after about two decades of the public internet, The Atlantic published a cover story asking "Is Google making us stupid?" Since then, and with the emergence of smart phones and social media, research has shown that too much time on the internet can lower our ability to concentrate, make us feel isolated and lower our self-esteem.
One 2011 review found that people increasingly turn to the internet for difficult questions and are less able to recall the information that they found on the internet when using it to answer those questions. Instead, participants had an enhanced ability to recall where they found it.
“The internet has become a primary form of external or transactive memory, where information is stored collectively outside ourselves,” the authors concluded.
People can certainly learn from artificial intelligence, but the danger lies in forgoing the learning process to simply regurgitate information that it feeds us.
In 2021, Fisher co-authored research that also found people who used internet searches more had an inflated sense of their own knowledge, reporting exaggerated claims about things they read on the internet compared to a control group who learned things without it. He termed this phenomenon the “Google effect.”
“What we seem to have a hard time doing is differentiating where our internally mastered knowledge stops and where the knowledge we can just look up but feels a lot like our knowledge begins,” Fisher said.
Many argue that AI takes this even further and cuts out a critical part of our imaginative process. In an opinion piece for Inside Higher Education, John Warner wrote that overrelying on ChatGPT for written tasks “risks derailing the important exploration of an idea that happens when we write.”
“This is particularly true in school contexts, when the learning that happens inside the student is far more important than the finished product they produce on a given assignment,” Warner wrote.
Much of the energy dedicated to understanding how AI affects our brains has been focused on adolescents because younger generations use these tools more and may also be more vulnerable to changes that occur because their brains are still developing. One 2023 study, for example, found junior high school students who used AI more had less of an ability to adapt to new social situations.
Another 2023 paper also found that students who more heavily relied on AI to answer multiple choice questions summarizing a reading excerpt scored lower than those who relied on their memory alone, said study author Qirui Ju, a researcher at Duke University.
“Writing things down is helping you to really understand the material,” Ju told Salon in a phone interview. “But if you replace that process with AI, even if you write higher quality stuff with less typos and more coherent sentences, it replaces the learning process so that the learning quality is lower.”
To get a better idea of what is happening with people’s brains when using large language models, researchers at the Massachusetts Institute of Technology connected 32-channel electroencephalograms to three groups of college-age students who were all answering the same writing prompts: One group used ChatGPT, another used Google and the third group simply used their own brains.
Although the study was small, with just 55 participants, its results suggest large language models could affect our memory, attention and creativity, said Nataliya Kos'myna, the leader of the “Your Brain on LLM” project, and a research scientist at the MIT Media Lab.
We need your help to stay independent
After writing the essay, 85% of the group using Google and the group using their brains could recall a quote from their writing, compared to only 20% of those who used large language models, Kos'myna said. Furthermore, 16% of people using AI said they didn’t even recognize their essay as their own after completing it, compared to 0% of students in the other group, she added.
Overall, there was less brain activity and interconnectivity in the group that used ChatGPT compared to the groups that used Google or their brains only. Specifically, activity in the regions of the brain corresponding to language processing, imagination and creative writing in students using large language models were reduced compared to students in other groups, Kos'myna said.
The research team also performed another analysis in which students first used their brains for the tasks before switching to performing the same task with the large language models, and vice versa.
Those who used their brains first and then went on to try their hand at the task with the assistance of AI appeared to perform better and had the aforementioned areas of their brains activated. But the same was not true for the group that used AI first and then went on to try it with just their brains, Kos'myna said.
“It looks like the large language models did not necessarily help you and provide any additional interconnectivity in the brain,” Kos'myna told Salon in a video call. “However, there is potential … that if you actually use your brain and then rework the task when being exposed to the tool, it might be beneficial.”
Whether AI hinders or promotes our capacity for learning may depend more on how we use it than whether we use it. In other words, it is not AI that is the problem, but our overreliance on it.
Van Belle, in Belgium, now uses large language models to write social media posts for his company because he doesn’t feel like that is where his skills are most refined and the process can be very time-consuming otherwise.
“I would like to think that I would be able to make a fairly decent LinkedIn post by myself, but it would take me an extra amount of time,” he said. “That is time that I don't want to waste on something I don't really care about.”
These days, he sees AI as a tool, which it can be — as long as we don't offload too much of our brain power on it.
“We’ve been on this steady march now for thousands of years and it feels like we are at the culmination of deciding what is left for us to know and for us to do,” Fisher said. “It raises real questions about how best to balance technology and get the most out of it without sacrificing these essentially human things.”
