Nataliya Kosmyna made an uncommon choice in a research lab at MIT’s Media Lab in Cambridge, Massachusetts: she published a study prior to peer review. She was afraid of what might occur if she waited, not because the results were not complete. Peer review can take up to eight months. She was concerned that during that time, a policymaker might decide that integrating AI into kindergarten curricula was a good idea. The study had to be published. Right now. “Developing brains are at the highest risk,” she stated bluntly.
Over the course of several months, 54 participants from the Boston area, ages 18 to 39, were split into three groups and asked to write SAT-style essays as part of a relatively small study. ChatGPT was utilized by one group. Google Search was used by one. There was no use at all. After attaching EEGs to each of them and tracking activity in 32 different brain regions, researchers observed what transpired. The outcomes weren’t subtle. In every metric the researchers monitored, the ChatGPT group displayed the least amount of brain engagement. Instead of getting better over time, they got worse. Many had practically stopped writing and begun copying by the time they finished their third essay. Give ChatGPT the prompt. Make a sentence better. “It was more like, ‘just give me the essay,'” Kosmyna remarked. The assignment was finished. Very little was discovered.
A different narrative was presented by the brain-only group. Particularly in the alpha, theta, and delta frequency bands linked to creativity, memory load, and semantic processing—the cognitive machinery that writing actually requires when you’re doing it yourself—they displayed the highest neural connectivity. They asserted that their work was their own. Their level of satisfaction was higher. They were more inquisitive. The ChatGPT output was largely described by two English teachers who evaluated the essays from all three groups as “soulless”—a term that carries subjective weight but also names something real: the lack of a thinking person behind the words.
The final section of the study is the most difficult to ignore. Participants were asked to rewrite one of their earlier essays after finishing them, but the ChatGPT group was required to do so without the aid. Almost nothing of what they had written remained in their memories. They avoided the deep memory processes that typically consolidate learning, as evidenced by the weaker alpha and theta activity in their brain wave patterns. “The task was executed, and you could say that it was efficient and convenient,” Kosmyna replied. “But you basically didn’t integrate any of it into your memory networks.” Eighty-three percent of ChatGPT users couldn’t remember what they had just written. Since the study’s publication, that figure has been making the rounds in education circles, and it should continue to do so.
The consequences for K–12 educational institutions are not hypothetical. They happen right away. According to Common Sense Media research, only a small percentage of parents are aware that over 50% of students between the ages of 12 and 18 use ChatGPT for academic purposes. Schools all over the nation are actively debating AI policies; some support its use, some forbid it, and many fall into an ambiguous middle ground that amounts to turning a blind eye. The MIT study does not support prohibition. Kosmyna specifically points out that the brain-only group performed well when later given access to ChatGPT, indicating that AI can supplement rather than replace a student’s ability to develop true cognitive ownership of a subject. The sequencing poses a risk. AI first, comprehension later is not a teaching approach. It is a quick fix that doesn’t require any thought.
This has been observed in clinical practice by Dr. Zishan Khan, a psychiatrist who treats children and teenagers. He observes kids who use AI extensively for schoolwork, and he explains the long-term cognitive risk in specific terms: when neural connections supporting memory retrieval, factual recall, and intellectual resilience are not used, they deteriorate. That is not a theoretical issue. It’s the way the brain functions. Additionally, it particularly affects developing brains, which are found in every K–12 school across the nation.
It’s difficult to ignore the fact that the technology sector has advanced far more quickly than research. November 2022 saw the public release of ChatGPT. Within months, schools started incorporating it. Two and a half years after the tool was introduced in classrooms, the MIT study was conducted in June 2025 without any systematic testing of its impact on the developing minds it was already reaching. In that case, Kosmyna’s choice to publish prior to peer review was not motivated by impatience. There was a pressing need. The outcomes are currently being taught in classrooms. All the research was doing was catching up.
