Your Brain on ChatGPT: A Cognitive Neuroscientist’s Breakdown

Since the launch of ChatGPT in 2022, many people have wondered about the long-term impact of generative AI on our ability to remember information, communicate through writing, and think critically. This June, a research group from MIT’s media lab published one of the first studies to address these questions in “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” 

One important thing about this paper, outside of the science, is that it’s a preprint. 

A preprint is a research article that scientists upload online before submitting it for formal peer review and publication by a scientific journal. This practice has become more popular in recent years because it allows researchers to get new and important research to the public quickly and without paywalls. However, it also means that the claims made in preprints should be viewed cautiously since the methods and results have yet to be approved by other experts in the field.

With this information in mind, let’s examine what this preprint says about generative AI’s impact on our brains.

Who and what are we looking at?

My initial questions when reading a paper like this are all about how the study was conducted.  How many people did they study? What kind of information did they get from study participants? How much information did they get? Did they use tests and tools that are appropriate for the questions that they are asking? 

Typically, all this information is found in a section titled “Methods and Materials” in the middle of or at the end of the paper. However, in this paper, the methods are found under “Experimental Design.” 

According to this section, this study took 54 people from various Boston universities and divided them into three different groups: a ChatGPT group, SearchEngine group, and a Brain-Only group. Each group has 20 minutes to write an essay based on an SAT prompt. Participants wrote three different essays on three separate occasions over the course of four months, but with slightly different rules:

• The ChatGPT group could use ChatGPT to assist with writing, but no other tools.
• The SearchEngine group was only allowed to use a standard search engine, like Google.
• The Brain-Only group couldn’t use any tools.

Researchers used an electroencephalogram (EEG) device to record each participant’s brain activity while they were writing.

Are there any notable limitations to their methods?

The first thing I noticed is the number of people studied, which is also known as the sample size. 54 total participants means that they had about 18 people in each group which is below average for the field. EEG studies typically average 21 people per group and the field is generally pushing to increase that number. 

Additionally, essay writing requires people to use a wide variety of mental skills, but there are still many other mental skills that people don’t engage with when writing. These other skills are not being assessed here and may be impacted differently by generative AI use. The authors do acknowledge some of these limitations in their “Limitations and Future Work” section.

So, what did they find?

The results of this paper are broken down by different analysis methods. The results discussed here are largely from the “EEG Analysis” section. 

People who wrote essays using only their brains activated a wide range of brain regions with strong communication between regions. The active brain regions were largely those responsible for more complex skills like memory, language, and shifting attention between tasks.

The SearchEngine group had less overall activity and communication between brain regions. However, people in this group had higher activity in parts of the brain responsible for vision-related functions, likely due to the visual aspect of searching for information online. 

People in the ChatGPT group activated the least number of brain regions and had weaker communication between regions compared to the other two groups. Researchers noted that the most active brain regions in this group were mainly related to organizing information and physical movement. 

These results suggest that writing without external tools requires people to use many parts of the brain, particularly those responsible for complex skills, and promotes communication between brain regions. On the other hand, using ChatGPT while writing essays allows people to complete the same task with less brain activity in fewer brain regions and with less communication between regions. The researchers behind this study refer to this phenomenon as “cognitive offloading.” People outsource their mental work to AI and therefore don’t engage parts of the brain typically needed to complete tasks. 

This could be considered a good thing! Most people want to preserve their mental energy at some points. However, the brain is a muscle and like the muscles in your leg or arm, the brain can become weaker without use.

What can we conclude from this?

Is it possible that constantly offloading mental work to AI could damage our cognitive abilities in the long term? 

Unfortunately, this study takes place over the course of four months, which isn’t enough time to truly answer this question. We will need more studies like this in the future to get definitive answers. However, this study sheds light on the fact that we may truly be using our brains differently when we use generative AI to get tasks done. It is a good start to addressing people’s questions about the impact of generative AI on brain function. This paper also serves as a well-timed reminder that expending the mental effort to do more complicated tasks can be beneficial in the long run.