Your Brain on ChatGPT: The MIT Study Every Parent Should Know

Note: I’ve been thinking and writing about AI’s use with your child’s thinking skills since 2023, particularly the implications and role of LLMs in education. I’ve highlighted the sparse research and championed cautious optimism in using it as a learning tool, while emphasizing the need for essential thinking skills. The 2025 MIT study in this post has greatly informed my recommendations for parents.

The purpose of Thriving Little Thinkers is to provide parents with low cost, high impact ideas for raising thinkers. With that goal in mind, I want to shine a light on how AI in education can come at an incredibly high cost: the cost of your child’s cognitive capacity.

The integration of AI into education is far outpacing our research and understanding of its effects. But in June 2025, researchers at MIT released results of a study that sparked viral attention: “Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task.” It’s the most helpful AI research TO DATE and I want to tell you all about it because it uses primary brain data, not just secondary subjective self report measures, to see what is happening in student brains when using AI to write essays.

Academic research investigating AI in education (specifically large language models or LLMs), is a brand new area of study. There is very little research, though it is growing, and much of the research is from self-reported survey answers or secondary analyses (looking for trends in data you already have, such as GPA, test scores, course grades, etc.). The bulk of the research is focused on assessment based outcomes, NOT the cognitive processes that are required to produce those grades and GPAs.

The MIT study is special, because instead of racing to find a positive outcome from using AI in education, it is the first study we have of an objective look at the COST of using LLMs.

The team released the results on their website instead of waiting for the full peer review process because they are very concerned with the current push for AI in K-12 classrooms. By getting their study results out as soon as possible, they aim to help inform educational administrative decisions.

As a parent navigating AI tools with your children, you might be wondering, what does this research actually mean for raising a thriving thinker?

This post shares what parents need to know about this study:

What the Study Measured
What the Study Found (& DIDN’T find)
Why Does This Matter?
Parent Takeaways
Conclusion & References

1. What the Study Measured

Researchers at MIT’s Media Lab wanted to understand what happens in the human brain when people use ChatGPT for writing tasks (compared to no tools at all or using internet searching). They recruited 54 adults (ages 18-39) from universities in the Boston area and divided them into three groups:

LLM Group: Used ChatGPT to help write essays

Search Engine Group: Used Google to research and write essays

Brain-Only Group: Wrote essays with no external tools

Over four months, participants wrote three essays using their assigned tool (or lack thereof). Researchers measured brain activity using EEGs(electroencephalography), a device which tracks electrical activity across different regions of the brain. If your child or someone you know has had a seizure and needed to wear a special cap that monitors their brain’s electrical activity–this is the same type of device. In addition to analyzing the brain activity, the researchers also analyzed the quality, content, and characteristics of the essays themselves and interviewed participants about their experience.

Here’s the big twist…in a fourth session, the groups switched. The ChatGPT users had to write without any tools, and the brain-only group got to use ChatGPT.

The choice of a writing task for this study is significant because writing is one of the best assessments we have for thinking. It allows us to “capture” a complex internal mental process and assess it externally. It requires managing multiple cognitive tasks at the same time (remembering and retrieving information, organizing it, synthesizing it, maintaining focus, considering the audience). How someone handles this cognitive load reveals the quality of their executive functioning (the capacity of their brain to act like the conductor of an orchestra and coordinate all the different regions and roles of the brain toward a specific goal). Writing is also one of the easiest things to outsource to an LLM.

2. What the Study Found

The results revealed measurable differences between the groups at three levels: neural (brain activity), linguistic (writing characteristics), and behavioral (memory and ownership).

Brain Connectivity: The brain-only group SHINED. They had the strongest and most widespread brain connectivity during essay writing. The search engine group showed moderate connectivity. The ChatGPT group displayed the weakest brain connectivity. Brain connectivity essentially measures communication happening between different regions of the brain (more connectivity suggests more cognitive engagement with the task).
- When your brain is in active learning mode (such as reading or writing), you utilize multiple regions and areas of the brain–simultaneously– with incredible processing power. We would expect to see an engaged brain with cross connectivity across regions, hemispheres, and lobes of the brain when someone is in learning mode. We would expect to see the front of the brain, particularly the prefrontal cortex, activated as the “conductor” of all this activity as it guides and directs the different areas to accomplish the chosen goal.
- Unfortunately, that’s not what was observed in the brain activity of the LLM-assisted group.

Memory and Ownership: When asked to recall and quote from their own essays just minutes after writing them, the ChatGPT group struggled and could rarely communicate what they had just submitted. They also reported the lowest sense of ownership over their work. In contrast, the brain-only group felt strong ownership,and could quote easily from their work. The search engine group fell in between.
- Narration has been described as an active process of making information your own. Narrating something you have written or read helps cement ideas, strengthens connections, and organizes new information into your internal knowledge structure. If students using LLMs to write can’t narrate or recall the information, they will not get the full benefits of the writing exercise. It’s also logical that they couldn’t recall the information if they didn’t generate the writing themselves.
- Given the concerns of plagiarism with LLM assisted writing, it is interesting that students using LLMs reported less “ownership” over the essays. Is it possible that less perceived ownership lends to less perceived accountability for the LLM-assisted essays?

Writing Patterns: The essays from the ChatGPT group were statistically similar to each other. They essentially repeated the same words, language, and ideas within topics.This suggests that when people rely on AI suggestions, similar writing is produced rather than developing unique perspectives. When teachers were asked to grade the essays, they easily picked out the LLM assisted essays vs. brain-only essays. When asked what made those essays stand out, the teachers described the LLM assessed essays as “almost perfect” but “soulless.”
- The MIT team conducted in-depth linguistic analyses that you can find here.

The Switch: When the brain-only group got to use ChatGPT in session 4, their brain connectivity actually increased significantly across all frequency bands. This suggests they were actively engaging with the AI tool rather than passively accepting its output. However, when the ChatGPT group had to write without AI assistance, they showed weaker connectivity and struggled more than the brain-only group had during their first session—even though both groups were essay-writing for the first time in that condition.

What This Study Does NOT Say…

It’s important to know the limits of any research study.

The sample was small and specific and did not include children. However, if the brain activity of young adults was impacted by LLM use, and K-12 students are in a “prime time” of brain development, I think it is logical and reasonable to ask “How much MORE so should we be concerned about the cost of LLMs in young students?”
The study only looked at essay writing, not using LLMs for other educational classroom tasks. It also didn’t break the writing tasks into components (idea generation vs. drafting vs. editing, etc.). It’s possible that using LLMs for some components of writing could have different benefits or costs than others.
Short Time Frame: Four months is not long enough to understand long-term impacts. When the brain-only group finally used ChatGPT, they showed increased brain connectivity. However, we don’t know if this benefit will stay with the students. It’s possible that if the brain-only group continues to use LLMs for writing support they will eventually have decreased connectivity, but we don’t have data past the 4th session. We don’t have a recommended “dosage” or cut-off point at which AI use becomes “overreliance” or at which we start to see decreased connectivity.
One AI Tool: The study used ChatGPT specifically. Different LLMs might produce different results.
Not Peer-Reviewed: This study has not yet been peer reviewed or through publication in a scientific journal. The researchers deliberately released this study early because they believe the topic is pressing and time-sensitive, especially as K-12 schools and colleges rapidly integrate AI tools into classrooms. The public scrutiny and commentary gained from an early release doesn’t replace a scrutinous peer review process, but it does point out large flaws more quickly. So far, other than the few limitations listed, I have not read any strong methodological arguments against this study.

3. Why Does This Matter?

Cognitive Debt. The researchers describe the results using the term “cognitive debt”—the idea that you’re “borrowing” on LLM’s efficiency now, but you’ll “pay” for it later. Relying on AI tools can reduce the deep cognitive activation that is necessary for independent thinking. Each time you outsource thinking to LLMs your brain practices less independent thought. Your neural networks for deep processing grow weaker, and your sense of ownership and accountability for the work you produce with LLMs declines. Over time and with repeated LLM use, these small deficits accumulate.

What starts as convenient assistance can become dependent thinking. The person who repeatedly uses LLMs for writing may find that when they need to write something important without it—during an exam, in a meeting, in a high-stakes situation—they struggle more than they would have if they’d worked their “brain-only” muscles.

The Brain Needs to Work: Just like muscles need resistance to grow stronger, our brains need cognitive challenge to build neural pathways, memory, and deep learning. When LLM tools do too much of the cognitive work, our children miss out on essential brain exercise. Struggle and effort are essential features of learning.

Imagine taking a forklift into the gym with you. You’re operating the controls to make the forklift lift the weight. Did the weights get lifted? Yes. Did you benefit from it, though? Was the purpose to grow stronger or to get the weight lifted? Without the necessary friction and mental effort that comes with deeply engaging with information, the product will reflect LLMs “thinking” capacity and not your child’s thinking capacity.

Developing Brains Are at Highest Risk: Dr. Kosmyna noted that younger learners with developing brains are “potentially in bigger danger” because their brains are just beginning to learn how to learn. The habits and patterns they form now will shape their cognitive abilities for life. Young brains (infancy through middle school) experience massive neural proliferation in which they build connections and circuits across the brain. Then during the middle school and high school years they continue to make connections and circuits but the front of their brain, where executive control is based, coordinates all of their faculties to work together. The middle and high school years also include neural pruning–what doesn’t get used gets lost. Those circuits aren’t maintained and your brain doesn’t spend energy on them. This timing in neural development makes the outsourcing of mental work all the more concerning.

Memory Requires Encoding: When participants used ChatGPT, the task was efficient and convenient, but they didn’t integrate (encode) information into their memory networks. For children who are still building foundational knowledge and skills, this presents a major concern. Encoding of the information is necessary for learning to happen not just in academics but in everyday life. Encoding information, integrating it into what you already know or don’t know, is a key difference between passive consumption and the formative nature of deep learning. LLM users were supervising an external source of knowledge, not building up their own internal knowledge base.

4. Practical Takeaways for Parents

Delay LLM integration until middle school and high school. Students need to develop strong internal neural networks, conceptual knowledge structures, and resilience in academic work BEFORE integrating AI as an educational tool.
Preserve brain-only experiences (or “normal” educational experiences as we would have called them 3 years ago): A significant portion of your child’s cognitive work should happen in “brain-only” mode, where they’re fully responsible for the thinking, the struggle, and the breakthrough. Intentionally create opportunities for your child to work without AI assistance. The brain needs practice doing hard cognitive work. Let them struggle with a math problem. Let them revise their own essay or their friend’s essay. Let them research a topic by reading actual books, even if they are developmentally capable of using AI tools.
Model AI use (and lack thereof). Remember that you are your child’s most influential model. If they only see you reaching for AI assistance at the first sign of difficulty, that’s the pattern they’ll internalize. Model the struggle of brain-only experiences AND model the intentional use of discernment when working with LLMs. If they see you mindlessly copying AI outputs, they’ll do the same. If they see you actively questioning AI, evaluating outputs, asking questions, and considering your own internal knowledge base as well–they’ll use that approach too. Your actions and behaviors should communicate to your children: “My brain is capable. Thinking is valuable. Struggle is productive. I don’t need AI or LLMs for every cognitive task.”
Teach information literacy.
- Require source checking and validation of LLM outputs. Look for hallucinations.
- Examine point of view and algorithmic bias. Is it “true” or is it a reflection of the LLM training data?

5. Conclusion

Sometimes the harder path, the one that requires more brain connectivity, more memory engagement, more cognitive effort, is exactly the path our children need to take. We need much more research in this area, but we can’t always wait for a big popular research study before we act. LLMs are already here, in our schools, and in our children’s hands. The speed of AI integration into education is unprecedented. As parents, it is our job to be informed and intentional. We need to preserve cognitive challenges that build strong brain networks and sound minds while also preparing our children for a world where AI collaboration is a reality.

Let’s not sacrifice our child’s prime time for neural development at the altar of convenience and efficiency. Build their brains now and incorporate the bots in education, later.

For more on this topic, check out my other posts on AI and education:

What are your thoughts on AI use in your child’s education? How are you balancing the benefits and concerns? I’d love to hear from you in the comments below.