Does ChatGPT Cause Cognitive Decline? What the MIT “Brain on ChatGPT” Study Actually Found

No, the widely shared MIT study does not show that ChatGPT causes a “massive decline in cognitive function,” brain damage, or a 47% loss of brain power.

It did find something worth taking seriously.

In a small experiment involving 54 mostly university-affiliated adults, people assigned to use ChatGPT while writing short essays generally showed less distributed EEG connectivity than people writing without outside help. They also had substantially more difficulty recalling exact language from their essays, particularly during their first session. Researchers interpreted the pattern as evidence that outsourcing parts of the writing process may reduce some forms of cognitive engagement. (DOI)

But that is a much narrower finding than the version now circulating on social media.

The experiment did not measure a 47% decline in overall cognition. It did not demonstrate falling IQ, impaired executive function, permanent memory loss or deterioration of the brain. It did not follow heavy ChatGPT users continuously for four months and discover that their minds progressively weakened.

And, importantly, the researchers themselves have explicitly asked journalists not to describe their results using terms such as “dumb,” “brain rot,” “damage” or “brain damage.” MIT still describes the work as preliminary. (MIT Media Lab)

The more defensible conclusion is both less dramatic and more useful:

Using AI to perform cognitive work for you can reduce how much of that particular cognitive work you perform yourself. Whether repeated reliance eventually weakens transferable skills remains an important research question—but this MIT experiment did not establish that it does.

The viral claim did not come from MIT

The wording in the widely circulated image can be traced to a June 21, 2025 post by author Brad Stulberg titled A New Study Shows ChatGPT is Dramatically Weakening Our Brains.

Stulberg described the MIT research as showing a “massive decline in cognitive function” from over-reliance on ChatGPT. That characterization is his interpretation of the research, not a finding stated by MIT researchers. (Brad Stulberg)

That distinction matters because the MIT team subsequently published unusually explicit guidance about how its findings should be described.

Asked whether it was accurate to say LLMs are making people “dumber,” the researchers answered no. Their FAQ specifically warns against language implying brain damage, brain rot or that LLMs simply make people stop thinking. (MIT Media Lab)

This does not mean the researchers think their results are meaningless. Quite the opposite: their paper raises concerns about what they call “cognitive debt” when people repeatedly delegate intellectual work to AI.

But “possible cognitive costs during AI-assisted writing” is not synonymous with “massive cognitive decline.”

What the MIT researchers actually did

The study, Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, was initially released in June 2025 and substantially revised in December 2025. The current preprint is 216 pages long. (DOI)

Researchers initially recruited 60 adults. Fifty-five completed at least three sessions, although the analysis reported 54 so the subjects could be divided evenly among three groups. Participants were between 18 and 39 years old, with an average age of about 23, and were recruited from MIT, Wellesley, Harvard, Tufts and Northeastern. (Benoît Labourdette production)

The 54 analyzed participants were randomized into three groups of 18:

They wrote SAT-style essays during three separate experimental sessions while wearing a 32-channel EEG system.

Each writing period lasted only about 20 minutes. (C2ST)

A smaller fourth session was later conducted with 18 participants. Nine previous ChatGPT users wrote without ChatGPT, while nine previous brain-only writers were allowed to use it.

Researchers compared EEG connectivity patterns, characteristics of the essays, participant interviews, essay scores, recall and perceived authorship.

That is an interesting experimental design.

It is also a very specific one.

The study investigated how adults’ brains and behavior differed while completing a constrained essay-writing task with or without AI assistance. It did not administer a comprehensive before-and-after battery of tests for general intelligence, memory, attention and executive function and find those abilities had deteriorated.

That difference is central to understanding the study.

No, ChatGPT users did not “lose 47% of their brain connectivity”

One of the most viral claims surrounding the research is that ChatGPT caused a 47% reduction in brain connectivity.

That number appears to have been calculated from one particular result in the paper.

In an alpha-frequency EEG comparison, the researchers counted 79 statistically significant directional connections favoring the brain-only group and 42 favoring the LLM group. Because 42 is approximately 47% lower than 79, online posts began describing this as a “47% reduction in neural connections” or even a 47% loss of brain activity. (Fermat’s Library)

That is not what the numbers mean.

The researchers were analyzing a measure called the dynamic Direct Transfer Function, or dDTF, which estimates directed statistical relationships between signals recorded at different EEG electrodes.

The brain-only participants did not literally possess 79 neural connections while ChatGPT users possessed 42.

Nor did researchers determine that 47% of the ChatGPT users’ brains had stopped functioning.

A later methodological commentary makes this distinction particularly clear: a “significant connection” in the analysis represents a statistical coupling that was significantly stronger in one experimental group than another. Counting fewer such differences does not necessarily mean weaker overall neural activity. (arXiv)

This is probably the single most important correction to the viral version of the story.

79 versus 42 is a feature of a particular statistical comparison of EEG signals. It is not a measurement of somebody losing 47% of their brain’s processing capacity.

The memory result is real—and genuinely interesting

The behavioral evidence is harder to dismiss.

Immediately after the first writing session, 15 of the 18 ChatGPT participants—83.3%—failed the study’s initial quotation measure, compared with two of 18 participants in each of the search and brain-only groups. The paper also reports that none of the ChatGPT participants produced a correct quotation under its stricter quotation-correctness measure in that first session. (Fermat’s Library)

That is a striking difference.

It is also narrower than saying 83% of ChatGPT users “lost their memory.”

Participants were being asked to reproduce language from an essay they had just created with varying amounts of outside assistance. If ChatGPT generated significant portions of the wording, poorer verbatim recall is not especially mysterious: those participants had not necessarily generated and encoded every sentence themselves.

The measure may still capture something important about cognitive engagement. But remembering an exact sentence is not equivalent to measuring general memory ability, conceptual understanding or long-term learning.

A methodological commentary on the study specifically questioned whether reproducing a sentence is a sufficiently strong proxy for the kinds of learning and memory processes the paper discusses. (arXiv)

There is another important detail that rarely appears in viral summaries: the enormous quotation gap shrank in later sessions.

By session two, only two of 18 LLM users reported being unable to provide a quotation at all, although four still failed the paper’s correctness measure. The participants now knew what they were going to be asked afterward, which itself changes the cognitive task. (Fermat’s Library)

So the first-session result is legitimate and potentially important.

It is not evidence that 83% of ChatGPT users became unable to remember what they were doing.

“The study lasted four months” is also easy to misunderstand

Another common description says researchers studied people using ChatGPT “for four months.”

Technically, the experimental sessions were conducted across approximately four months.

But this was not a four-month trial in which researchers assigned one group to rely heavily on ChatGPT throughout everyday life and then watched their cognitive abilities deteriorate.

Participants completed three controlled essay-writing sessions during that period, with an optional fourth session for some subjects. (Peter Attia MD)

That is a considerably smaller exposure.

The difference is crucial if the question is whether long-term habitual AI use causes lasting cognitive decline.

This experiment cannot answer that question.

What happened when the ChatGPT users stopped using ChatGPT?

The fourth session is arguably the most provocative part of the research.

Eighteen participants returned. Nine who had previously written with ChatGPT were now required to write without it, while nine former brain-only participants were given ChatGPT.

The former ChatGPT group continued to show differences in some neural and behavioral measures after the tool was removed. Researchers interpreted this as possible evidence of accumulated cognitive effects.

But this result should be treated as especially preliminary.

There were only nine people in each crossover condition. Participants were also writing about topics they had encountered previously, introducing possible familiarity and practice effects.

The methodological commentary argues that without appropriate same-condition control groups for the fourth session, practice effects and other confounders make strong causal interpretation difficult. (arXiv)

Even the revised paper contains a more complicated picture than the viral summaries suggest. Some measures remained reduced, while others showed substantial cognitive effort once former ChatGPT users had to work independently.

That is interesting evidence about task adaptation.

It is not evidence that ChatGPT caused lasting brain damage.

There are legitimate methodological questions about the study

The MIT experiment is ambitious. Combining EEG recordings, writing analysis, interviews and crossover conditions can generate hypotheses that simpler studies cannot.

But ambitious designs also create statistical problems.

A December 2025 methodological commentary by researchers from the University of Vienna and Technische Universität Dresden identified several concerns, including the small sample, unclear reporting, reproducibility issues and interpretation of the EEG analysis. (arXiv)

Among the more important issues:

The original analysis included just 18 people per main condition and nine per condition in session four. The commentary’s illustrative power analysis estimated that roughly 159 participants would be needed under the assumptions it selected for a repeated-measures design.

The EEG analysis also involved a very large number of electrode-pair comparisons. The paper describes conducting up to approximately 1,000 repeated-measures analyses, while the commentators argued that the reporting around multiple-comparison correction was not sufficiently clear to independently evaluate all of the results.

The commentary further noted that the paper said 55 people completed the experiment but analyzed 54, without clearly explaining the exclusion criterion beyond creating equal groups. (arXiv)

None of those criticisms prove the MIT findings are wrong.

They do mean the most dramatic numerical claims should not be treated as settled neuroscience.

And there is a temporal wrinkle worth noting: the critique was submitted on December 29, 2025, while MIT’s substantially revised version was posted December 31. Some reporting changed in the revision, so the critique should not be treated as a line-by-line review of every aspect of the current 216-page version.

The underlying concerns about sample size, interpretation and the enormous analytic search space nevertheless remain relevant.

The study still does not appear to be peer reviewed

As of September 2026, I could not find a peer-reviewed journal version of the MIT study.

The current arXiv record lists the December 31, 2025 revision as version two. MIT’s own project page continues to characterize the research as a preprint, and a peer-reviewed July 2026 article discussing the study likewise refers to it as a preprint. (DOI)

That does not make the research worthless. Preprints are a normal part of modern scientific communication.

It does mean the research has not yet received the additional methodological scrutiny implied when social-media posts simply call something an “MIT study” and present its interpretation as established fact.

Newer research makes the picture more complicated, not less

The most important reason not to dismiss the underlying concern is that other research is finding evidence of cognitive offloading when AI replaces effortful learning.

But the broader literature does not support a simple rule that “ChatGPT makes your brain worse.”

A 2025 randomized controlled trial involving 120 undergraduates compared students who used ChatGPT while learning AI concepts with students using traditional study methods. On a surprise test 45 days later, the ChatGPT group scored 57.5%, compared with 68.5% in the traditional-learning group. The authors interpreted the result as evidence that unrestricted AI assistance can undermine durable memory by reducing productive cognitive effort. (ScienceDirect)

That result is more directly relevant to long-term retention than remembering an exact essay sentence several minutes later.

Meanwhile, a 2025 peer-reviewed CHI study surveyed 319 knowledge workers and analyzed 936 real-world examples of generative-AI use. Greater confidence in the AI was associated with less self-reported critical-thinking effort, while greater confidence in one’s own expertise was associated with more. The study also found that AI changed the type of thinking people performed—from generating information and solutions toward verifying, integrating and supervising AI output. (DOI)

Again, that supports an over-reliance concern.

But it was a survey of self-reported behavior, not evidence that AI caused cognitive deterioration.

Then consider the evidence pointing in the other direction.

A 2026 meta-analysis pooled 35 experimental studies involving 4,193 participants and found a moderately positive overall effect of ChatGPT interventions on student learning outcomes, with an effect size of g = 0.670. Results varied meaningfully with the subject, duration and way ChatGPT was integrated into instruction. (Nature)

That meta-analysis has its own limitations. The underlying studies were conducted between 2022 and 2024, the interventions were heterogeneous, and positive classroom outcomes do not answer whether habitual outsourcing might weaken particular skills over many years.

But it makes one point difficult to avoid:

The scientific literature does not support treating “using ChatGPT” as a single cognitive exposure with a universally harmful effect.

A 2026 EEG experiment reveals why “less brain activity = worse” is too simple

One particularly useful study was published in Acta Psychologica in August 2026.

Researchers randomized 34 undergraduates to learn a digital-image-processing concept either through ChatGPT or from a human instructor while recording EEG activity.

The ChatGPT group achieved comparable basic retention but performed worse when asked to transfer what they had learned to new situations.

Yet unlike the MIT essay study, the ChatGPT group showed higher theta activity across much of the brain during learning. (ScienceDirect)

That is an important reality check.

If more EEG activity automatically meant healthier, deeper or more effective cognition, the ChatGPT group should have learned better.

It did not.

The researchers instead interpreted the elevated activity as potentially reflecting greater visually driven cognitive effort from interacting through text, while the human instructor provided more efficient multimodal scaffolding.

In other words:

Brain activation is not a scoreboard where a larger number automatically means better thinking.

That is precisely why transforming MIT’s 79-versus-42 connectivity result into “47% less brain function” is scientifically indefensible.

So can ChatGPT become a cognitive crutch?

Yes. That is plausible, and there is now evidence supporting the concern.

But the mechanism matters.

If a student needs practice constructing an argument and asks ChatGPT to construct the argument, the tool has removed precisely the effort the student was supposed to practice.

If someone is learning to code and repeatedly asks AI to solve every debugging problem before trying to diagnose it, that person may become very good at evaluating generated solutions while failing to develop the same independent debugging ability.

If a writer has AI generate the ideas, outline, prose and revisions, it would be surprising if that writer encoded the resulting material as deeply as someone who created it personally.

That is cognitive offloading.

Humans have always done it. Writing offloads memory. Calculators offload arithmetic. GPS offloads navigation. Search engines offload factual retrieval.

Offloading is not inherently pathological. Sometimes offloading is the entire purpose of a tool.

The real question is whether you are outsourcing an activity whose cognitive process you actually want to retain.

That distinction is almost completely absent from claims that AI simply is “bad for your brain.”

Should AI really “not be your first line of defense for basic tasks”?

Not as a universal rule.

If your objective is learning, immediately asking AI for the finished answer can eliminate productive struggle that helps create durable knowledge. The research increasingly supports caution there.

But if your objective is simply getting a low-value task completed, there is no obvious cognitive virtue in deliberately making yourself perform every operation manually.

Someone who lets software alphabetize a spreadsheet has not suffered cognitive decline because they could have sorted the rows themselves.

The better rule is:

Do not outsource the part of a task that contains the skill you are trying to develop or preserve.

For learning, that may mean attempting the problem first, retrieving what you know from memory, constructing your own explanation and then using AI to critique it, identify weaknesses, challenge assumptions or generate additional practice.

For routine production work, using AI immediately may be entirely rational.

The scientific evidence currently supports that distinction much better than a blanket “don’t use AI for basic tasks.”

What would it take to show that ChatGPT actually causes cognitive decline?

A convincing test would look considerably different from the MIT essay experiment.

Researchers would need a much larger randomized sample, baseline measurements of memory, attention, reasoning and executive function, controlled or accurately logged AI exposure over an extended period, comparison groups performing equivalent work without AI, repeated validated cognitive testing, and follow-up after AI assistance was removed.

Ideally, the study would also distinguish very different styles of AI use: asking for a complete answer, brainstorming collaboratively, receiving Socratic questions, checking completed work, tutoring, summarizing and simple clerical automation.

That distinction may ultimately prove more important than the binary question of whether somebody “uses AI.”

Until those studies exist, claims that ChatGPT causes generalized cognitive deterioration run ahead of the evidence.

And what about the claim that AI is “awful for the environment”?

That is a separate issue from the MIT cognition study.

AI infrastructure has a real and rapidly growing environmental footprint, particularly because large data centers consume electricity and require cooling, equipment and grid infrastructure.

The International Energy Agency currently projects global data-center electricity consumption to rise from roughly 485 terawatt-hours in 2025 to around 950 TWh in 2030, with electricity use from AI-focused data centers growing even faster. (IEA)

At the same time, the environmental conversation has also become more complicated than the viral “every ChatGPT query is environmentally disastrous” framing.

The IEA reports that the electricity required for an individual AI task has fallen extremely rapidly as hardware and software have become more efficient. Simple text queries are relatively inexpensive, while advanced reasoning, agentic workflows and video generation can require hundreds or thousands of times more energy per task. Total electricity use can therefore rise dramatically even as individual operations become more efficient, because AI adoption and computational intensity are growing even faster. (IEA)

So there are two simultaneously true facts:

AI’s aggregate energy demand is becoming significant enough to affect electricity systems.

And environmental impact varies enormously depending on the model, data center, electricity source and type of AI task.

Calling AI simply “awful for the environment” is an opinion. The defensible factual claim is that AI has material environmental costs, those costs are growing rapidly at the infrastructure level, and the footprint per use varies dramatically.

None of that was tested by the MIT brain study.

What the MIT study actually tells us

There is a real warning buried underneath the sensationalism.

When ChatGPT performs more of an essay-writing task, people appear to perform less of some of the cognitive work associated with generating that essay themselves. In MIT’s experiment, that difference was visible both in EEG patterns and in immediate recall.

Other research suggests that unrestricted AI assistance can sometimes hurt long-term retention or knowledge transfer.

But other experiments and meta-analyses find neutral or positive learning effects when AI is integrated differently.

The emerging evidence therefore points toward something more useful than “AI makes you stupid”:

AI can either replace thinking or support thinking, and those are not cognitively equivalent uses of the technology.

The MIT study provides preliminary evidence for that distinction.

It does not show a massive decline in cognitive function.

It does not show that users lost 47% of their brain connectivity.

It does not establish brain damage.

And it does not prove that using ChatGPT for ordinary tasks will make a person’s cognitive abilities deteriorate over time.

Those conclusions would require evidence we do not yet have.

The most important unanswered question is no longer whether humans will offload cognitive work to AI. We plainly will.

It is which kinds of cognitive work we can safely outsource—and which kinds we need to keep doing ourselves if we want to remain good at them.

References and Further Reading

Original MIT research

• Kosmyna et al. — “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task” (arXiv, revised December 2025) — The current 216-page version of the original study and the primary source for its experimental findings.

• MIT Media Lab — Your Brain on ChatGPT project overview and researcher FAQ — MIT’s explanation of the experiment, limitations and explicit warning against describing the findings as brain damage or proof that AI makes people “dumber.”

Methodological scrutiny and interpretation

• Stanković et al. — Commentary on “Your Brain on ChatGPT” — Detailed methodological critique addressing sample size, EEG analysis, statistical reporting, reproducibility and interpretation of “significant connections.”

• Pereira Campos and Koff — “Your brain on ChatGPT, but whose brain?” (Frontiers in Developmental Psychology, 2026) — Peer-reviewed discussion of the MIT findings and the limitations of translating reduced neural connectivity directly into poorer cognition.

Broader evidence on AI, learning and cognitive offloading

• Barcaui — “ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention” — Randomized experiment finding lower 45-day retention among undergraduates assigned to unrestricted ChatGPT-assisted learning.

• Lee et al. — “The Impact of Generative AI on Critical Thinking” (CHI 2025) — Peer-reviewed study of 319 knowledge workers showing how generative AI changes critical-thinking effort and shifts work toward verification and supervision.

• Wu et al. — “ChatGPT’s impact on student learning outcomes: a meta-analysis of 35 experimental studies” (2026) — Meta-analysis of 4,193 participants finding an overall positive learning effect while showing that outcomes depend strongly on instructional design and context.

• Zhang et al. — “Exploring ChatGPT’s potential in dialogic teaching: A comparative neurocognitive study of AI and human instruction” (Acta Psychologica, 2026) — Randomized EEG study finding similar retention but weaker knowledge transfer with ChatGPT, despite greater activity in some EEG frequency bands.

Source of the viral framing

• Brad Stulberg — “A New Study Shows ChatGPT is Dramatically Weakening Our Brains” — The June 2025 commentary containing the “massive decline in cognitive function” framing circulated in the screenshot; this language is not MIT’s conclusion.

AI energy use

• International Energy Agency — Key Questions on Energy and AI — Current IEA analysis of AI-related data-center electricity growth, rapid per-task efficiency improvements and the enormous differences in energy use between simple text generation and more computationally intensive AI workloads.

Editorial currency note: This article reflects research and publication status checked through September 3, 2026. The MIT study remains an evolving preprint, and the evidence base on generative AI and cognition is developing rapidly.