Editorâs Note: Peter Simons did not use an LLM or any other AI in the creation of this article, either to summarize the study or to generate the text of this piece.
A new study out of MIT has found that those who used large language models (LLMs) like ChatGPT to write an essay had significantly lower brain activation than those who didnât. They were also unable to remember anything they had âwrittenâ just minutes ago. While it was easier to write an essay this way, the text also regurgitated the same talking points and lacked creativity and critical thinking.
Those who used LLMs were essentially turning off the creative, thinking parts of their brains, and thus they did not learn, remember, or feel when composing the essay.
And it had a lingering effect: when switched to having no LLM assistance, those who had previously used LLMs still fared poorly. Rather than activating the creative, critical thinking parts of their brains when forced to put away ChatGPT, their brain activation mostly showed that they tried to remember what the LLMs had written previously.
According to the researchers, this has profound implications for education, creativity, and intellectual development. LLMs are already pervading our daily life, replacing artists, teachers, and writers. This may contribute to the fading away of the uniquely human ability to create something new or to think through a problem in unconventional ways.
The study, called âYour Brain on ChatGPT,â was posted before peer review on preprint server arXiv. It was led by MIT Media Lab research scientist Nataliya Kosâmyna.

âInteraction with AI systems may lead to diminished prospects for independent problem-solving and critical thinking. This cognitive offloading phenomenon raises concerns about the long-term implications for human intellectual development and autonomy,â the researchers write.
They add, âThis shift from active information seeking to passive consumption of AI-generated content can have profound implications for how current and future generations process and evaluate information.â
Educational outcomes have already been slipping in the developed world; a third of US 8th graders are unable to read at basic level. And while itâs true for kids, itâs true for adults tooâmore than half of US adults read below a 6th-grade level.
This trend began even before the pandemic, though that may have worsened it. Part of it could also be due to smartphonesâcountries that have banned smartphones in schools have seen better learning outcomes for their students, and the research supports the notion that just having a phone near you is detrimental to your attention and focus.
So, maybe AI is the last big leap that will take us beyond the point of no returnâbut in which direction?
Although the results of Kosâmynaâs study were striking, it was a small study in a highly educated population and focused on only one task, essay writing with a strict time limit. So this study alone canât answer these massive questions. Yet there is, of course, other evidence we can look at to make our determination.
So, is the use of AI going to enhanceâor hinderâthe human raceâs ability to think?
AI as an Accessibility Aid?
Proponents argue that using artificial intelligence (AI) in education evens the playing field, helping those who are worse at intellectual pursuits perform better. Indeed, tools like ChatGPT and other AI aids are more likely to be used by those who are already struggling, while those who had confidence in their ability to write and learn barely use them.
But the MIT study suggests that using AI in this way actually further weakens the abilities of those who already struggle, diminishing their capacity to learn and think critically in the futureâhelping them only to fall further behind.
Moreover, since LLM use led in this study to more generic and repetitive essays lacking in personal expression, it seems that LLMs are erasing the unique voices of those who use them, rather than amplifying their ability to express themselves.
If the goal is for differently-abled people to be able to tell their stories in their own unique voices, or to be able to argue coherently for better support, then they might be in trouble, because LLMs are having the opposite effect. Itâll help you complete an essay, sureâbut that essay will be generic and superficial.
The researchers write, âWhen individuals fail to critically engage with a subject, their writing might become biased and superficial. This pattern reflects the accumulation of cognitive debt, a condition in which repeated reliance on external systems like LLMs replaces the effortful cognitive processes required for independent thinking.â
Worse, LLM users are also more susceptible to manipulation and propaganda, according to the researchers, as their brains become out of practice at thinking critically. Those who use AI become increasingly accepting of what they are told and less willing to exert their own thought processes.
As the researchers write, âCognitive debt defers mental effort in the short term but results in long-term costs, such as diminished critical inquiry, increased vulnerability to manipulation, decreased creativity. When participants reproduce suggestions without evaluating their accuracy or relevance, they not only forfeit ownership of the ideas but also risk internalizing shallow or biased perspectives.â
Proponents may argue that the gain of efficiently creating new work outweighs the concerns that differently-abled people might fall behind. But the newly-created content can also have significant issuesâissues that those who outsource their thinking to AI might not even be able to detect.
For instance, âvibe codingâ has become a popular trend, with companies seeking to hire people who do not themselves know how to code, but who know how to prompt an LLM to program a code for them. Proponents argue that this means web resources can be built much more efficiently, saving a company time and money. But the outcomes are concerning: vibe coders donât understand the work they are doing, and so they leave security vulnerabilities for hackers to exploitâholes in their code that they donât even know exist, much less how to plug.
So, those vibe coders who created a program that interfaces with your bank to automate debt collection? That could lead to the biggest financial meltdown weâve ever seen.
And the implications for education, as seen in the MIT study, are even more striking.
LLMs vs. Search Engines vs. Brains
Kosâmynaâs study included 55 participants, all taken from colleges in the Boston area, including MIT as well as Wellesley, Harvard, Tufts, and Northeastern. Thirty-five were undergraduates, 14 were âpostgraduateâ students, and 6 had obtained an MSc or PhD and were working at their institution. The majority (32) were female.
Fifty-four of the participants were assigned into three equal groups:
- LLMs: Those who were told to use ChatGPT. They were not allowed to use anything else (like search engines or other apps).
- Search engines: Those who were allowed to use any website, but explicitly prevented from using LLMs. Google was their default browser. All search engines had â-aiâ added to all searches so that no LLM would be used.
- Brain-only: Those who were not allowed to use LLMs, websites (including search engines), or other tools.
There were three regular sessions, one per month. In each of the sessions, the participants were tasked with writing an essay within 20 minutes. The prompts came from the infamous college admissions test, the SAT. Each session, participants were offered three prompts to choose from.
There was also a fourth session, in which 18 people participated. In session four, the researchers switched which group the participants were in, so those who originally used LLMs were only allowed to use their brains, while those who had not used any aids were now instructed to use LLMs to write their essays.
The participants were fitted with an EEG headset to measure brain activity during the essay writing task (it was calibrated to each participant right before the test).
After the essay writing task was finished, the researchers conducted a short interview with the participants, including asking them to provide a quote from their essay and asking them about their sense of ownership: how much they feel the essay was theirs.
Finally, the researchers conducted a Natural Language Processing (NLP) analysis to make sense of the content of the essaysâwas the content of the essays noticeably different based on which group the participants were in?
Results: âFocused on Copying and Pasting Contentâ
The main finding was that brain activity was strongest in those who did not use any aids, weaker in the search engine group, and weakest in the LLM group.
âWriting an essay without assistance (Brain-only group) led to stronger neural connectivity across all frequency bands measured, with particularly large increases in the theta and high-alpha bands,â the researchers write.
In addition, the type of brain activity was different, with those who used search engines activating their brain circuits to process visual information and organize it, but not entering the deep, rich, interconnected cognitive state of those who used no aids.
Essentially, those who did not use any aids activated circuits across their whole brain, in an interconnected way, to accomplish all the various tasks that were required: creativity, language, organization, planning, critical thinking, working memory, etc. Those who used search engines offloaded a little of this cognitive work and thus activated less of the brain, with a focus in different areas.
âSearch Engine group’s lower alpha connectivity (and fewer frontal hubs) might reflect reduced reliance on internal memory due to the availability of online information, consistent with the âGoogle effect,â wherein easy access to external information can diminish the brain’s tendency to internally store and connect information,â the researchers write.
And those who used LLMs barely activated their brains at all.
âThe directed connectivity analysis reveals a clear pattern,â the researchers write. âWriting without assistance increased brain network interactions across multiple frequency bands, engaging higher cognitive load, stronger executive control, and deeper creative processing. Writing with AI assistance, in contrast, reduces overall neural connectivity, and shifts the dynamics of information flow. In practical terms, a LLM might free up mental resources and make the task feel easier, yet the brain of the user of the LLM might not go as deeply into the rich associative processes that unassisted creative writing entails.â
But how does this lack of brain activity impact memory, learning, and thinking ability?
One way of checking that is to see whether the participants remembered their work immediately after writing the essay.
Just minutes after the first session, nobody in the LLM group was able to correctly provide a quote from their essay. Yet almost everyone in the other two groups was able to correctly quote their work (15/18 in the search engine group and 16/18 in the brain-only group). This suggests that the LLM group did not internalize anything they had âwrittenâ since they were mostly just copying and pasting AI-generated text.
âThe complete absence of correct quoting in the LLM group during Session 1, and persistent impairments in later sessions, suggested that not only was memory encoding shallow, but the semantic content itself may not have been fully internalized,â the researchers write.
The interesting thing, though, is that many of those in the LLM group thought they were writing. Half of those who used an LLM (9/18) stated that they had âfull ownershipâ of the essay and some stated that they were only using the LLM for spelling and grammarâdespite not being able to remember a single thing they supposedly wrote.
The implication here is not just that people who use LLMs arenât doing the workâitâs that they think theyâre doing the work even when they are not.
However, some people were able to admit that they were mostly just copying and pasting the LLMâs generated text, although they made excuses for it: “I tried quoting correctly, but the lack of time made it hard to really fully get into what ChatGPT generated,” said one participant.
âMost of them focused on reusing the tools’ output, therefore staying focused on copying and pasting content, rather than incorporating their own original thoughts and editing those with their own perspectives and their own experiences,â the researchers write.
This was borne out by the NLP analysis, which found that the LLM group efficiently produced essays, but that the writing was âgeneric,â with âweaker argumentationâ than those who used their brains.
Session Four: âSkill Atrophyâ
During session four, the participants were provided with three prompts to choose fromâthe three they had already written about in the other sessions.
One striking result: Two-thirds of those who had used LLMs for the first three sessions did not even recognize the prompts, despite having (supposedly) written an essay on each. All of those who had used their brains, of course, did indeed recognize the prompts as essay topics they had already written on.
Another result: Those who had used LLMs for three sessions (and now were in the brain-only group in session four), found their brains a little more active than those in the brain-only group in session one (unpracticed), but less active than the brain-only group in sessions two and three (after having practiced using their brains for the task).
âThis interpretation is supported by reports on cognitive offloading to AI: reliance on AI systems can lead to a passive approach and diminished activation of critical thinking skills when the person later performs tasks alone,â the researchers write.
âThis resonates with findings that frequent AI tool users often bypass deeper engagement with material, leading to âskill atrophyâ in tasks like brainstorming and problem-solving,â they add.
Those who used their brains for three months, then were forced to use LLMs, had increased brain activity while using the LLM. The researchers write that this could be due to having to integrate this novel stimulus into a task that the participants were already practiced in doing. After all, they write, âTaskâswitching studies show that shifting from one rule set to a novel one reâexpands connectivity until a new routine is mastered.â
That is, even if session four LLM use did activate more brain circuits, it could just be due to extra work having to integrate the new user interface, not leading to any worthwhile improvements.
The researchers did not investigate what would have happened if those who practiced using their brains continued to use only their brains in session four. Thus, it is impossible to say whether LLM use after using your brain and then an AI is actually beneficial, or whether you should just keep using your brain.
âLess Sophisticated Reasoning and Weaker Argumentationâ
The findings of this study indicate that those in the LLM group experienced lower cognitive loadâthey had to devote less of their mental energy to composing and writing the essay. The researchers make clear that this is correlated with efficiency (you can produce more text), but it is also correlated with worse thinking (your arguments are lower-quality):
âWhile lower cognitive loads often improve productivity by simplifying task completion, LLM users generally engage less deeply with the material, compromising the germane cognitive load necessary for building and automating robust schemas. Students relying on LLMs for scientific inquiries produced lower-quality reasoning than those using traditional search engines, as the latter required more active cognitive processing to integrate diverse sources of information,â they write.
And the type of brain activity shifts too, from using your brain to think critically to using it to take in information in a copy-and-paste manner.
The researchers write, âThe reduction of cognitive load leads to a shift from active critical reasoning to passive oversight.â
The difference here is between short-term passive engagement, emblematic of ChatGPT use, and long-term intellectual engagement, emblematic of brain use. For cognitive engagement, traditional methods are superior:
âWhile tools like ChatGPT and multi-role LLM are adept at fostering immediate and short-term engagement, there are limitations in maintaining intrinsic motivation over time. There is also a lack of deep cognitive engagement, which often translates into less sophisticated reasoning and weaker argumentation. Traditional methods tend to foster higher-order thinking skills, encouraging students to practice critical analysis and integration of complex ideas.â
In sum, using LLMs leads to what is called âmetacognitive laziness,â in which students take the easy path of generating AI responses rather than putting in the brain work required to really think through a complex topic. But without doing that brain work, students donât actually learn or even remember the material.
âAI tools that generate essays without prompting students to reflect or revise can make it easier for students to avoid the intellectual effort required to internalize key concepts, which is crucial for long-term learning and knowledge transfer,â the researchers write.
âSuch an interpretation aligns with concerns that over-reliance on AI can erode critical thinking and problem-solving skills,â the researchers add.
The findings of the NLP analysis demonstrate that using search engines and LLMs led to more âgenericâ essays, while using your own brain to write led to essays presenting unique perspectives.
The researchers had two English teachers read and evaluate the essays. Hereâs their feedback about them:
âSome essays across all topics stood out because of a close to perfect use of language and structure while simultaneously failing to give personal insights or clear statements. These, often lengthy, essays included standard ideas, reoccurring typical formulations and statements, which made the use of AI in the writing process rather obvious. We, as English teachers, perceived these essays as ‘soulless’, in a way, as many sentences were empty with regard to content and essays lacked personal nuances. While the essays sounded academic and often developed a topic more in-depth than others, we valued individuality and creativity over objective âperfectionâ.â
The Echo Chamber
The implications for education are striking: AI use can efficiently create content, but it does not translate to actual learning. Instead, LLM use leads to the erosion of memory and critical thinking ability.
âThis trade-off highlights an important educational concern: AI tools, while valuable for supporting performance, may unintentionally hinder deep cognitive processing, retention, and authentic engagement with written material. If users rely heavily on AI tools, they may achieve superficial fluency but fail to internalize the knowledge or feel a sense of ownership over it.â
The researchers express concern that using these aids, especially LLMs, lead to repetitive, generic work. In essence, those who use these aids to write are reproducing the same arguments that have already been made rather than thinking through their own perspectives on them.
âThe LLM group produced statistically homogeneous essays within each topic, showing significantly less deviation compared to the other groups. The Search Engine group was likely, at least in part, influenced by the content that was promoted and optimized by a search engine,â the researchers write.
Worse, those who use these aids are not being exposed to counterpoints, disagreements, or other information that might lead to either changing an opinion or being able to more clearly defend a point of view. Instead, these aids exacerbate the âecho chamberâ effect. LLMs are the worst offender here, as they are programmed to sycophantically agree with and praise the user for their âinsightâ rather than debate.
âThe implications for educational environments are particularly significant, as echo chambers can fundamentally compromise the development of critical thinking skills that form the foundation of quality academic discourse,â the researchers write.
Moreover, since LLMs have a conversational style and will often cite a number of sources (even if âhallucinatedâ), they may mislead the user into thinking theyâre receiving various views when they are actually receiving sycophancy.
âWhen students rely on search systems or language models that systematically filter information to align with their existing viewpoints, they might miss opportunities to engage with challenging perspectives that would strengthen their analytical capabilities and broaden their intellectual horizons. Furthermore, the sophisticated nature of these algorithmic biases means that a lot of users often remain unaware of the information gaps in their research, leading to overconfident conclusions based on incomplete evidence,â the researchers write.
Ultimately, this may erode humanityâs ability to engage in independent reasoning:
âThis creates a cascade effect where poorly informed arguments become normalized in academic and other settings, ultimately degrading the standards of scholarly debate and undermining the educational mission of fostering independent, evidence-based reasoning,â the researchers write.
The Promise and the Peril
One interpretation of the MIT study that has bounced around social media is that using AI after doing your own work first somehow increases your intelligence and critical thinking. Essentially, proponents of AI believe that it has positive effects as long as you use it after using your own mind to its fullest extent. For instance, medical doctor Yudara Kularathne writes on social media: âUsed wisely, it makes you superhuman.â
Yet the study certainly doesnât imply this. Being generous, this conclusion could be a misinterpretation of the EEG data in the fourth session, from the group that switched from brain-only to LLMs. This group had more brain activity when using the LLMs, but the researchers did not test whether just using your brain a fourth time would be better.
They also acknowledge that the additional brain activity here could be because these participants had to learn a whole new interface (they were confronted with the AI after not having used it). This brain activity could be an indicator of novel stimuli, not of improvements in thinking.
This study was not set up at all to answer the question of whether using your brain first, and then an AI, was better than just using your brain consistently.
But there is one hint: those who switched to using LLMs after using their brains for three sessions did not continue using their original ideas. Instead, according to the researchers, they âgave in to LLM suggestionsâ and created the same generic, AI-generated essay as those who had only used LLMs. The researchers suggest that âparticipants may have leaned on the model’s suggested phrasing with relatively little further revision.â
This means that despite having done this task three times with their human brains alone, when given an AI assistant in the fourth session the participants just accepted its generated text. They did not somehow use it to become âsuperhumanâ and write an even better essay. They used it to write a worse, more generic essay more efficiently.
So, anyone on social media drawing a hyperbolic conclusion that AI âmakes you superhumanâ if you just use it differently is peddling nonsense. That claim is absolutely not supported by the study.
Is AI âLike a Calculatorâ?
AI proponents describe it as being like a calculatorâan obviously helpful tool that arguably has had no negative effects. It makes the process of doing math so much easier that we now canât imagine life without it, but there is little concern about its effects on human development.
But is a calculator actually a good metaphor for AI?
For one thing, a calculator will give you the correct answer every time, which makes it pretty different from an LLM. After all, LLMs have become notorious for generating falsehoods, including fake citations and made-up studies in incredibly important contexts like government healthcare reports and legal settings like courtrooms. They exacerbate biases, which can be especially problematic in healthcare and law. AI-powered chatbots are even now being sold as therapists to millions, all while encouraging delusional thinking and helping plan suicide.
Has a calculator ever encouraged you to commit suicide? Has a calculator ever convinced someone that they are are a spiritual guru who has âaccessed the secrets of the universeâ and landed them in a psychiatric hospital? Has a calculator ever even made up a fake answer to the math equation you typed into it?
Yet AI provides these lies entirely convincingly, all while praising the user in whatâs been called âsycophancy.â
This is the danger: Not just that AI is being used to offload our cognitive abilities so that our brains donât have to work. The danger is that itâs convincing us that we are not offloading our cognitive abilities in this way.
You can see it in the language of the proponents, as in Kularathneâs Tweet: âUsed passively, it makes you dull. Used wisely, it makes you superhuman.â
AI proponents believe that they use AI âcorrectlyââas a way to augment their existing thinking. But the results of the MIT study contradict this interpretation. They show that even if you use AI after already doing work, you are prone to copy-paste its generic, weak work rather than continuing to critically evaluate and improve your own work.
And most frightening of all is that you will continue to believe you did the work yourselfâdespite not being able to quote a single line you have supposedly written just minutes afterward. Youâll think itâs supercharging your intellect, turning you into a spiritual genius who can see beyond the illusion, like Neo in The Matrix. But what youâll actually be doing is generating generic, vague, poorly argued content at a much faster rate.
It doesnât matter whether youâre a vibe coder who doesnât know there are holes that hackers are using to steal your userâs data, or a lawyer citing made-up cases to support your client, or a healthcare admin putting together a list of fake references to support the governmentâs policies, or a therapy provider encouraging your clients to commit suicide. In all of these cases, you will think you are becoming superhuman, but you will actually be ruining human lives.
And the AI companies themselves know this. Theyâve done the studies themselves and theyâre posted on the internet for all to see. They just donât care. Their goal is to make the entire world so dependent on AI that we canât live without it. An infinite customer base, forever. And the best strategy? Start âem young, just like with cigarettes.
The more we outsource our thinking to AI, the more we erode our own ability to critically think. The more we erode our ability to critically think, the more we have to outsource our thinking to AI. Itâs this recursive loop that forms the basis of Microsoft, OpenAI, and Anthropicâs wet dreams. If we give in to the loop, we will let Idiocracy become reality.
**