In June 2025, MIT released a study titled, Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.
The Study
Researchers at the MIT Media Lab conducted this study to understand the “cognitive cost” of using artificial intelligence tools like ChatGPT for writing essays. They tracked 54 students from universities like MIT and Harvard who were asked to write SAT® essays.
These students were divided into three groups: one using only their brains, one using Google search, and one using ChatGPT. To see what was happening inside their heads, every participant wore an electroencephalography (EEG) headset that recorded their brainwaves in real time.
This study was driven by the question of whether using a generative AI tool like ChatGPT makes us better writers or if we are just “offloading” our intelligence to the machine. It’s important to note that since the research paper has not yet been peer-reviewed, the research team advises that “all the conclusions are to be treated with caution and as preliminary.”
Conclusions
After finishing their study, the researchers found a striking pattern. The more help that a student had when writing their essays, the quieter their brain became. While the “Brain-only” group showed strong, widespread neural activity, the ChatGPT users had the weakest overall brain connectivity. This suggests that the AI was doing the heavy mental lifting, leading to a trade-off where using a tool for a quick result today results in weakened critical thinking and memory skills later. The researchers refer to this phenomenon as “cognitive debt.”
The effects were also visible in the students’ behavior and in the quality of their work. Just minutes after finishing their essays, the ChatGPT users were generally unable to quote even a single sentence they had just written. They also reported feeling less ownership over their essays, often viewing the machine as the true author.
When experienced human teachers graded the essays without knowing which group wrote them, they described the AI-assisted work as “soulless.” While these essays were often grammatically perfect and well-structured, they lacked the personal insights and unique voice found in the work of students who wrote without AI.
Ultimately, the study suggests that while AI can be a helpful assistant, using it as a shortcut may prevent the brain from engaging in the productive struggle necessary for deep learning.
Takeaways
Here are nine takeaways derived from the study’s conclusions:
1. Neural Activity Scales With Effort: Brain connectivity is strongest and most widespread when writing without any digital tools. Conversely, using a large language model (LLM) like ChatGPT elicits the weakest overall neural coupling, suggesting that the machine is handling the heavy mental lifting.
2. The “Memory Gap”: Reliance on AI significantly impairs a person’s ability to recall what they just wrote. In the study, 83% of LLM users could not provide a single correct quote from their own essay minutes after finishing, whereas nearly 90% of those in the Brain-only and Search Engine groups could.
3. Fragmented Sense of Ownership: Students using AI reported a diminished sense of authorship and agency. While Brain-only participants claimed full ownership of their work, LLM users often felt like partial authors, viewing the AI as the primary creator.
4. Accumulation of “Cognitive Debt”: The researchers found that repeated reliance on AI creates “cognitive debt,” a trade-off where using a tool for immediate convenience today results in a long-term decline in independent critical thinking and creativity.
5. Technically Perfect but “Soulless”: While they were grammatically correct and well-structured, essays written by AI lacked the unique personal insights and individual voice found in essays written by the Brain-only group. Brain-only writing showed significantly more variety and unique stylistic choices.
6. Strategic Timing Is Critical: The study suggests that AI should only be introduced after a student has engaged in self-driven cognitive effort. Participants who wrote without AI tools first and used AI later—the “Brain-to-LLM” group—showed much higher neural engagement and used the tool more strategically than those who first started with AI.
7. Shift in Brain Role: Using AI shifts the brain’s activity from generating content to merely supervising it.
8. Bypassing Productive Struggle: While AI reduces immediate cognitive load, which makes the task feel easier, it also bypasses the deep analytical processes required to internalize knowledge and build robust mental schemas. Again, lack of productive struggle leads to “cognitive debt.”
9. Brain-to-LLM Wins: While participants who used AI to write their initial essays experienced cognitive debt, the opposite was true for students who turned to AI after writing their first essays without it. When these students eventually turned to AI, they experienced a spike in brain connectivity, suggesting that they were more engaged and were actively reconciling the AI’s suggestions with their own internally stored plans. Their essays also scored above average, and they demonstrated better integration of content compared to their previous “Brain-only” sessions. This group wrote better prompts and maintained high memory recall presumably because they had already engaged in the productive struggle.
Implications for Teachers
At a practical level, this study suggests that teachers should plan lessons that mitigate the “cognitive debt” that can be introduced by using generative AI too early or during critical stages of the learning process. To that end, included below are six ways teachers might design lessons that ensure students continue to build robust neural networks for learning.
1. Prioritize the brain-first sequence.
The study found that the order in which tools are introduced matters significantly. Participants who wrote without tools first and used AI later—the “Brain-to-LLM” group—showed much higher neural engagement and used the tool more strategically than those who started with AI from day one.
Therefore, teachers should design lesson plans that require a brainstorming or first-drafting phase without technology use. Students should generate their own core ideas and structure before being allowed to use AI to refine or extend their thoughts.
2. Implement the quoting test for retention.
The memory gap for AI users in the study is striking, as 83% of LLM users could not quote a single sentence from their own work just minutes after finishing it, while Brain-only and Search Engine groups had near-perfect recall.
To assess whether deep learning has occurred, teachers might consider moving away from just grading the final written product and, instead, ask students to provide a memorized quote from their work or a 2-minute oral summary of their main arguments. If students cannot recall what they wrote, the AI likely did the thinking for the student.
3. Value individuality and unique voice over technical perfection.
AI-assisted work is often described as “soulless,” as it’s technically correct but lacks personal insight and a unique voice. AI writing is also statistically homogeneous, reusing the same structures and concepts.
In response to this, teachers can adjust their rubrics to prioritize personal anecdotes, unique stylistic choices, and creative deviations over perfect grammar and standard academic structure. They might even reward imperfections that show original human thought over the middle-of-the-road prose typically produced by LLMs.
4. Transition students from creators to supervisors.
The study noted that using AI shifts the brain’s role from generating content to supervising it. If this is true, teachers should make sure that students are consciously and actively participating in that supervision when using AI. Students can be required to track their changes and reflect on how they critiqued, filtered, and edited the AI’s suggestions as they worked through the writing process. This will be an increasingly important skill as AI use grows.
5. Create productive struggle zones.
In MIT’s study, neural connectivity was strongest when students had to struggle to find their own words and organize their own thoughts. When students used AI to bypass this struggle, they suffered from skill atrophy and weakened critical thinking.
In our classrooms, we need to make sure that students embrace struggle and understand that it’s through this struggle that they will grow. We should explain the science to students and help them realize that just as lifting weights builds muscle, the friction of trying to articulate a complex idea builds a stronger brain. As part of this approach, teachers might explicitly designate certain lessons as “Struggle Zones,” where digital tools are limited and the struggle is embraced.
6. Implement a human-to-AI-to-human progression.
The report suggests that the timing of AI introduction is critical. Withholding LLM tools during the early stages of a task may promote durable memory traces and robust neural networks, allowing students to leverage AI more effectively and autonomously later in the process.
The key is to create an AI sandwich, with humans using their natural brains first, bringing in AI once they have engaged in cognitive struggle, and then bringing the human brain back in at the end to assess the value of the AI’s input. This sequence of learning and engagement with AI leverages the best of the human experience, preserves cognitive struggle and growth, and also takes advantage of the benefits of AI.
AVID Connections
This resource connects with the following components of the AVID College and Career Readiness Framework:
- Instruction
- Rigorous Academic Preparedness
- Student Agency
- Insist on Rigor