Harvard Says AI May Be Dulling Our Minds. Here’s What the Data Actually Shows

Anyone who has used ChatGPT long enough is familiar with the moment when you sit down to write without it and the words just don’t come out the way they used to. The cursor is blinking. The page remains blank. It’s difficult to describe how something that was once effortless now feels laborious. It might be a diversion. It might be exhaustion. However, an increasing amount of research indicates that it may be more structural and more challenging to undo.

The Harvard Gazette published an article in November 2025 with the widely shared headline, “Is AI Dulling Our Minds?” A variety of Harvard cognitive scientists and education researchers were asked the question, but they were unable to provide a reassuring response. The Harvard Graduate School of Education researcher Tina Grotzer was characteristically exact about it. She claimed that students are using AI without knowing how it operates, and that misplaced trust in the results is just one aspect of the issue. The underlying problem is that no one is assisting youth in realizing the power of their own minds.

However, the most startling data came from MIT rather than Harvard. Nataliya Kosmyna and her colleagues at the MIT Media Lab recruited 54 participants between the ages of 18 and 39 and divided them into three groups. Three groups used ChatGPT, Google Search, and nothing at all to write essays. Throughout, they wore EEG headsets, which allowed the researchers to observe the real-time activity of 32 different brain regions. By academic standards, the findings were concerning. Out of the three groups, ChatGPT users displayed the least amount of neural engagement. There was a discernible decrease in their brain connectivity, or the extent to which various brain regions were interacting and working together. And with every session after that, it got worse. Many participants had completely stopped writing by the third essay, and they were just asking ChatGPT to create the work and sending it over.

One particular detail set those findings apart from the majority of AI criticism. 83% of ChatGPT users were unable to quote a single line when the researchers later asked them to recall the contents of the essays they had just written. It wasn’t that the essays were unmemorable, but rather that they had never given the material much thought. None of the information had been encoded because their brains had not been exposed to it long enough. The task was completed. There was no learning.

Kosmyna honestly explained why she published the work prior to peer review, which can take up to eight months. The speed at which institutional decisions were being made worried her. “I am afraid there will be some policymaker who decides, ‘let’s do GPT kindergarten,'” she replied. It’s a legitimate fear. She contended that the policy debate is advancing more quickly than the process of validating research, and that developing brains are most vulnerable.

Treating the MIT study as settled science would be premature because of its small sample size and lack of peer review. However, it is not an isolated entity. The use of AI tools and critical thinking test scores were consistently found to be negatively correlated, according to a study published in the journal Societies that polled 666 participants of all ages and educational backgrounds. The 17–25 age group, which has grown up with these tools and, in many cases, reached adulthood without ever developing the cognitive habits they are now offloading, saw the biggest impact. The opposite pattern was seen in older participants. Higher critical thinking scores, less reliance on AI, and—most importantly—neural pathways that were developed before the means to circumvent them were available.

This phenomenon is known as “cognitive offloading” in cognitive science. It’s not brand-new. Writing things down rather than memorizing them, using calculators instead of performing math, and using GPS instead of learning to navigate are just a few examples of how humans have always delegated mental tasks to external tools. However, each of those tools offloaded a particular, bounded function. AI is not like that. It relieves the burden of reasoning. Building an argument, assessing a source, and solving a problem are all processes that, when repeated, create the mental framework that enables people to think clearly when it matters.

It’s difficult to ignore the well-known study on London taxi drivers, whose hippocampi were noticeably bigger than those of bus drivers due to the physical alteration of their brains caused by decades of memorization of the city’s streets. Use it or lose it is more than just a catchphrase. It explains a biological process. Researchers at Harvard, MIT, and other universities are currently asking what happens when a generation uses its intellect less.

Another Harvard professor, David Deming, provided a frame worth considering. It is impossible to download the higher-order skills that are important in an AI world, such as how to collaborate with others, handle information, and make decisions in uncertain circumstances. They are acquired gradually by doing. A classroom that generates thinkers is not the same as one that generates outputs. Furthermore, it’s still unclear if the educational system has fully taken that distinction into account given how quickly it is integrating AI.

The unsettling question at the heart of all of this is not whether AI is helpful, which it obviously is, but rather whether the way most people use it is subtly sacrificing long-term potential for immediate convenience. Don’t use the tools, according to the research. It’s saying that the brain needs to be given something to do, just like any other muscle. Additionally, there is currently a growing temptation to give it nothing at all.