Cognitive Offloading: What the MIT ChatGPT Study Found
Cognitive Offloading: What the MIT ChatGPT Study Found
Cognitive offloading means using a tool or action to reduce the mental work a task requires. The MIT "Your Brain on ChatGPT" preprint found that students who wrote essays with ChatGPT showed the weakest EEG connectivity of three groups, and in the first session most could not quote what they had just written. The study has not been peer reviewed. It did not show brain damage.
But it surfaced a real cost: when you let a tool generate and you just select, you skip encoding. You never really learn the thing you submitted. Here is what the paper measured, what it cannot tell you, and what cognitive offloading research says about using AI without losing the thinking that makes your work yours.
Key Takeaways
- Cognitive offloading is normal and often useful. Humans have offloaded memory to lists, calendars, and calculators for centuries. The question is what you offload.
- The MIT preprint (54 participants, EEG, 20-minute essays) found weaker brain connectivity in the ChatGPT group, but the study is small, limited to one task, and still under review as of October 2026.
- In session 1, 83.3% of ChatGPT users could not correctly quote their own essay minutes after writing it. Only 11.1% in each of the other two groups failed.
- The fix is not "stop using AI." It is sequencing: draft first, then bring in the tool.
The Study at a Glance
| Detail | What the preprint reports |
|---|---|
| Participants | 54 adults (ages 18 to 39, mean 22.9) from MIT, Wellesley, Harvard, Tufts, Northeastern |
| Task | 20-minute essays on SAT-style prompts, 3 sessions over 4 months, optional 4th crossover |
| Groups | ChatGPT only / Search engine (AI answers blocked) / Brain only |
| Measurement | 32-electrode EEG connectivity (dDTF), NLP essay analysis, teacher and AI-judge scoring, interviews |
| Headline result | Brain-only strongest connectivity; search intermediate; ChatGPT weakest |
| Session 4 sample | n = 18 (crossover: LLM-to-Brain and Brain-to-LLM) |
| Status | arXiv preprint, v1 June 2025, v2 Dec 2025, marked "under review." Not a peer-reviewed journal article as of October 7, 2026. |
| What it cannot tell you | Workplace use, other AI models, long-term outcomes, deep brain structures |
What Is Cognitive Offloading?
The term comes from a 2016 paper by Risko and Gilbert in Trends in Cognitive Sciences. They defined it as "the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand." Their examples: tilting your head to perceive a rotated image, programming a phone reminder instead of holding a meeting time in working memory.
Offloading is ancient. Grocery lists, wall calendars, the abacus. Sparrow's 2011 "Google effects on memory" study showed the same principle: when people expected future access to information, they remembered where to find it instead of the information itself. (Caveat: one of that paper's experiments did not replicate in the 2018 Social Sciences Replication Project.)
The trade-off is clear. You gain capacity for the next task. You lose encoding of the thing you handed off. Fine for formatting and calendar math. Expensive when the thing you offloaded is the argument your name is on.
What Did the MIT "Your Brain on ChatGPT" Study Actually Measure?
Fifty-four participants wore a 32-electrode EEG headset (Enobio, 500 Hz) and wrote essays on SAT-style prompts within 20 minutes. Three groups, three sessions over four months, plus a fourth crossover session (n = 18). EEG connectivity tracks how strongly brain regions coordinate during a task, not intelligence, and is not a "brain scan" in the MRI sense.
The reported pattern: Brain-only participants showed the strongest, most distributed networks. Search engine users fell in the middle. ChatGPT users showed the weakest connectivity. In session 4, LLM-to-Brain participants showed reduced alpha and beta connectivity, suggesting under-engagement even after the tool was removed. Brain-to-LLM users showed higher memory recall and activation of occipito-parietal and prefrontal areas.
When people had done the thinking first and then switched to ChatGPT, they stayed more engaged.
The Result That Matters Most: People Could Not Quote Their Own Essays
After each session, participants were asked to provide a correct quotation from the essay they had just written. In the ChatGPT group, 83.3% (15 of 18) failed in session 1. In both the search engine and brain-only groups, only 11.1% (2 of 18) failed. By session 3, 6 of 18 ChatGPT users were still failing.
Self-reported ownership was lowest in the ChatGPT group and highest in the brain-only group.
When the tool generates and you select, you skip the act of composing. Composing is how you encode. You can edit an AI draft for an hour and still not know what the document says. You cannot defend in a meeting what you never wrote.
Is AI Making Us Dumber? What the Study Cannot Tell You
The authors were explicit about limits: 54 participants (18 in crossover) from one geographic area, ChatGPT only, an essay task in an educational setting. Results "may not generalize across tasks." EEG cannot see deep structures like the hippocampus.
On their project FAQ, the authors asked media not to use words like "stupid," "dumb," "brain rot," "harm," or "damage." Lead author Nataliya Kosmyna told Tech & Learning: "We didn't find any brain rot."
Stankovic et al. (arXiv, December 2025) raised concerns about sample size, reproducibility, EEG methodology, and reporting inconsistencies. In Nature, Adam Green of Georgetown questioned the 18-person session 4, and Guido Makransky of the University of Copenhagen noted the setup does not reflect real-world AI use.
Honest verdict: a provocative pilot, not proof of lasting harm. Do not call it "debunked" either.
Cognitive Debt and AI Critical Thinking: What Other Studies Add
The MIT preprint is not the only data point. Two 2025 studies contribute, though both are correlational.
| Study | n | Design | Finding | Shows cause? |
|---|---|---|---|---|
| Gerlich 2025, Societies | 666 survey + 50 interviews | Cross-sectional survey | Negative correlation between frequent AI use and critical thinking, mediated by offloading; younger participants (17 to 25) scored lower | No |
| Lee et al., CHI 2025 (Microsoft Research + CMU) | 319 workers, 936 examples | Survey | Higher confidence in GenAI linked to less critical thinking; higher self-confidence linked to more | No |
The pattern across all three studies: the more you trust the tool to think, the less thinking you report doing. None prove AI causes cognitive decline. All of them should make you pause before hitting "generate" on work that carries your name.
A "Brain First, AI Second" Workflow for Knowledge Work
Practical inference from the study pattern, not a tested protocol.
- Write the first outline yourself, timed, no tools. Even 15 minutes of unassisted composition forces encoding. The Brain-to-LLM group stayed engaged because they drafted first.
- Ask AI to critique, not generate. "What's weak in this argument?" beats "Write me a draft." See these ChatGPT productivity hacks that save real hours.
- Offload what you do not need to own. Formatting, boilerplate, lookups. These are what offloading is for. More on how to use AI to be more productive without handing over your core reasoning.
- Before sending, close the chat and restate your key claim from memory. If you cannot, you do not own it.
- Keep one no-AI block per day for the work your role is judged on. Identify which of your tasks count as deep work. Bouncing between a chat window and your draft adds the real cost of context switching on top of the offloading problem.
Conclusion
The MIT preprint is a small, early study with real limits. Its 54 participants from Boston-area universities do not represent everyone who uses AI at work, and it has not been peer reviewed.
What it offers is a clean demonstration of an older principle: you remember and own what you build, not what you watch being built. Cognitive offloading is a feature of human cognition, not a flaw. But offloading the thinking rather than the logistics costs you the part that matters. Think first. Bring the tool in second. Test yourself before you call the work yours.
Frequently Asked Questions
Does ChatGPT make you dumber?
No single study can answer that. The MIT preprint found weaker EEG connectivity and poor quote recall in the ChatGPT group, but these are short-term observations of 54 participants during a 20-minute essay task. The authors asked people not to use words like "dumb" or "brain rot." What the data suggest is that generating text with ChatGPT reduces encoding, so you are less likely to remember or own what you submitted.
Is the MIT ChatGPT study peer reviewed?
No. As of October 7, 2026, it is an arXiv preprint marked "under review." The project site states that conclusions should "be treated with caution and as preliminary." A commentary by Stankovic et al. raised concerns about sample size, reproducibility, and EEG analysis.
What is cognitive debt?
In the preprint, cognitive debt refers to the accumulated cost of repeatedly offloading thinking to an AI tool: reduced brain connectivity, lower ownership, weaker recall across sessions. It is the paper's framing, not yet an established term in cognitive science.
Is cognitive offloading bad?
Not inherently. Writing a grocery list, setting a phone alarm, and using a calculator are all offloading. The cost shows up when you offload the part you need to learn or retain. Risko and Gilbert (2016) framed it as a trade-off between freeing cognitive resources and losing encoding.
How many people were in the MIT ChatGPT study?
Fifty-four adults completed sessions 1 through 3 (ages 18 to 39, from five Boston-area universities). Eighteen completed the fourth crossover session. The study recruited 60 initially.
Does using Google have the same effect as using ChatGPT?
In the MIT study, the search engine group sat between brain-only and ChatGPT on most measures. Their connectivity was weaker than brain-only but stronger than ChatGPT, and their quote recall was much better (only 11.1% failed). Sparrow's 2011 "Google effects on memory" found that search engines shift what you remember (location over content), though one of its experiments did not replicate in 2018.
For the No-AI First Draft, Bring Your Own Alertness
The study's clearest lesson: the draft you compose yourself is the one you remember and can defend. That first unassisted work block, often 20 to 60 minutes of hard composition, is the most effortful stretch of the day. It is also the stretch that matters most.
Roon is built for that block. Each pouch pairs 80 mg caffeine with 60 mg L-theanine, a combination linked to alertness and attention-switching accuracy in a systematic review of RCTs (Payne et al., Nutrition Reviews, 2025; confidence intervals showed uncertainty, and the review included industry-funded authors). It also contains 25 mg methylliberine (Dynamine) and 5 mg theacrine (TeaCrine) in a zero-nicotine, zero-sugar sublingual pouch. In a randomized, double-blind crossover PK study (n = 12), methylliberine co-administered with caffeine cut caffeine's oral clearance from 41.9 to 17.1 L/hr and extended its half-life from 7.2 to 15 hours with no increase in peak concentration (study). That study used 100 mg methylliberine plus 150 mg caffeine, roughly 4x Roon's methylliberine and 1.9x its caffeine, so the direction applies but the numbers are the study's, not Roon's. The felt window is 6 to 8 hours, onset in 5 to 10 minutes. Stop at least 8 hours before bed.
Roon does not make you smarter and does not improve memory. It keeps you alert for the work that requires your own thinking. From $9.16 per tin on subscription. Try it for that first draft.
Related from Roon
- ChatGPT Productivity Hacks
- How to Use AI to Be More Productive
- How to Improve Working Memory for Adults
- L-Theanine + Caffeine Task Switching: What the 2010 Trial Found
Written by Roon Team
