You might have seen the headlines: “Terrifying MIT Study Finds ChatGPT is Rotting Our Brains”,
or
Is ChatGPT making us dumb? MIT brain scans reveal alarming truth about AI’s impact on the human mind — The Economic Times
or social media posts that went viral, like this one:
All tributes to the recent MIT study, "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,"
Judging by all the AI-generated summaries in the newsfeed, I doubt most of them actually read the full study before hyping it up.
This is irony so pure, it deserves its own trophy.
A study on AI making us intellectually lazy goes viral, because AI made it easy to summarize without reading the full study, because AI made it so easy to spread an idea w/o anyone bothering to verify?
Yes, 206 pages is a slog.
Well… that’s why you subscribed, I guess ;) Do give yourself a pad.
I have been reviewing studies on how AI will impact human behaviour. I admit, the result of this study isn’t really a surprise, but I love how this study gives the world a hammer that AI is not almighty.
It is a great opportunity for us, group-pragmatic, to be loud and be heard.
Yes, I know. I just made fun of the viral posts. I do see that they serve a critical role in helping the general population become more aware of how AI can affect behaviour and the brain.
My previous work, like "Why Thinking Hurts After Using AI?" and "The Silent Classroom”, are two great complementary pieces to read if you’re interested in how using AI altered our personal and social behaviour.
This study, however, offers an as-close-as humans can get view of the mechanisms at play. Short of cutting your brain in half when using ChatGPT… nope, it wouldn’t work; you’d be gone by then.
They aimed to find out the answers to these questions:
Do people write significantly different essays when using LLMs, search engines, and with their brains alone?
How does your brain activity differ when using LLMs, search, or their brain-only?
How does using LLM impact your memory?
Does LLM usage impact your perception of ownership of your work?
TL;DR
Tool choice reshapes writing output and thinking. Essays and brain activity differ a lot depending on whether students used ChatGPT, Google (search only), or just their own brains.
ChatGPT-assisted writers produced strikingly homogeneous essays with repeated phrasing. Their brain scans showed up to a 55% drop in connectivity relative to brain-only writers. Only 1 in 9 students who used AI can correctly requote their work.
Only half (9/18) of students who use ChatGPT for their essays feel like they fully own the work.
The essays written by students who wrote without tools had the most diverse ideas and vocabulary. They also formed the strongest memory traces, demonstrated accurate recall, and felt proud ownership over their writing.
Another listen learned, Human vs AI “teachers“ saw things differently. Teachers graded the AI-crafted essays harshly for being formulaic and unoriginal. An automated AI “judge,” however, gave those same essays inflated scores.
This study has sparked numerous debates. Some argue that it lacks sufficient evidence, while others praise it because the graphic of a brain scan has all the elements needed for a study to go viral in comparison to previous AI impact studies on critical thinking.
I plan to cover this debate but also include the previous studies that I found valuable, not least this MIT one.
Shall we?
The Four-Month Experiment.
Over a semester-long study, researchers at MIT Media Lab followed 54 college students (aged 18–39) to see how AI assistance affects writing and learning.
Let’s say you were one of them.
The researchers split the participants among three groups:
🤖 one writing essays with ChatGPT (LLM group),
🔍 one using Google Search (search only, no access to LLM like Gemini),
🧠 and one with brain-only (no external tools). All students tackled similar SAT-style essay prompts under timed conditions (around 20 minutes per essay) across three sessions. You were in this one.
I have readers across more than 80 countries. So if you don't know what an SAT-style essay is, here are some examples.
Loyalty essay
Many people believe that loyalty whether to an individual, an organization, or a nation means unconditional and unquestioning support no matter what. To these people, the withdrawal of support is by definition a betrayal of loyalty. But doesn't true loyalty sometimes require us to be critical of those we are loyal to? If we see that they are doing something that we believe is wrong, doesn't true loyalty require us to speak up, even if we must be critical?
Assignment: Does true loyalty require unconditional support?
Happiness essay
From a young age, we are taught that we should pursue our own interests and goals in order to be happy. But society today places far too much value on individual success and achievement. In order to be truly happy, we must help others as well as ourselves. In fact, we can never be truly happy, no matter what we may achieve, unless our achievements benefit other people.
Assignment: Must our achievements benefit others in order to make us truly happy?
Crucially, in a fourth session, the researchers flipped the script for some (18 of them).
The ChatGPT-dependent group was then asked to write without any AI (🤖🤖🤖🧠), while the unplugged “brain-only” group got to use ChatGPT for the first time (🧠🧠🧠🤖).
Since you have been writing with your own brain, now you suddenly have access to and use AI to help you rethink some of your arguments.
This swap-round allows a before-and-after comparison within the same individuals. To make the most of the limited number of participants.
Throughout, you wore EEG headsets, and you’d look like this:
That recorded your brain activity across 32 regions as you wrote, and afterwards, you were asked about your writing process and asked to recall details from your work.
To assess writing quality, two English teachers graded the essays on factors like argument strength, originality, and style. The researchers also employed an AI judge to rate the essays.
This setup allows the team to observe both how the writing process and brain activities differ with AI assistance, as well as how human evaluators perceive the outcomes compared to those of an AI evaluator.
LLM, One Big Language Blender.
After the essays were collected, the texts were analyzed using natural language processing (NLP) to spot patterns in content and style.
The differences were quite obvious.
The ChatGPT-assisted essays were eerily uniform.
Human teachers closed the loop by detecting the LLM-generated essays, as they recognized the conventional structure and homogeneity of the delivered points for each essay within the topic and group.
In plain English, students in the LLM group often ended up using the same formulations and ideas as one another. The LLM group’s essays showed very little variation within each prompt.
Here's a comment from one of the English teachers:
These, often lengthy, essays included standard ideas, reoccurring typical formulations and statements, which made the use of AI in the writing process rather obvious…
By contrast, the brain-only group’s essays were far more diverse.
These students drew from their personal knowledge, experiences, and creativity, resulting in a wide range of perspectives and vocabulary in their writing. Their essays weren’t all polished perfection, but they were distinctively human, and each piece bore the author’s unique voice or approach.
Equally, students in the search-group essays were influenced by the same highly-ranked web results (for example, multiple search users latched onto the phrase “homeless person” when writing about philanthropy).
This is totally understandable.
Using Google led different students to fetch the same information, and thus gravitate toward similar content, much like the ChatGPT group. The ontology of topics covered by the search group overlapped significantly with that of the ChatGPT group, whereas the brain-only writers ventured into more varied territory.
So in short,
LLM group: Essays were highly homogeneous within each topic, showing little variation. …
Search Engine group: Essays were shaped by search-engine-optimized content; their ontology overlapped with the LLM group but not with the Brain-only group.
Brain-only group: Diverse and varied approaches across participants and topics…
All’ve been said!
The same English teacher complained about the homogeneity of the AI-generated content but still thinks that there are some advantages to using AI.
… However, some of these obviously AI generated essays did offer unique approaches, e.g. examples or quotes, which then led to higher uniqueness scores, even if structure and language lacked uniqueness.
Brain Engagement Plummets with AI
Here comes the part everyone's been talking about for the last two weeks.
The EEG frequency differences between AI and brain-only methods were dramatic.
Writing without any aid lit up the brain; writing with AI dimmed it.
In neuroscience terms, the brain-only group showed the strongest, most widespread neural connectivity during the task, engaging high-level thinking networks across the cortex.
In EEG frequency bands associated with creativity, memory, and attention (alpha, theta, delta waves), the brain-only writers led the pack.
Take the theta band for Happiness topic, for example (see the screenshot), it is thought to reflect broad, large-scale cortical integration and may relate to high-level attention and monitoring processes even during active tasks.
These two rows of images filter out the “noise” and show only the connections that are both statistically significant and strong.
Here’s how you read them:
Thin blue lines: Connections that are there, but weak (often just above the “statistical bar”).
Thick red lines: Connections that are both statistically significant and strong (the most robust brain communication).
So the upper row scan result shows the LLM group has many fewer or weaker lines, suggesting less brain engagement when using ChatGPT; the bottom brain scan with red lines in the Brain-only group, which means their brains had stronger, more coordinated activity during essay writing.
This MIT scans show: Writing with Google dims your brain by up to 48%. ChatGPT pulls the plug, 55% less neural connectivity.
Brain-to-LLM (🧠🧠🧠🤖): When Strong Brains Learn to Outsource
When Brain-only participants were introduced to ChatGPT in Session 4, they showed higher neural connectivity across all frequency bands compared to participants who had used LLM from the start.
Note, this represents reactivation of cognitive networks rather than enhanced performance.
The research found that Brain-to-LLM students demonstrated 'higher memory recall, and re‑engagement of widespread occipito-parietal and prefrontal nodes', suggesting their prior unaided writing helped them maintain cognitive engagement even when using AI tools.
This means that the sequence of tool introduction matters!
Brains that develop independent thinking capabilities first may be better positioned to use AI as a genuine assistant rather than a cognitive crutch.
LLM-to-Brain (🤖🤖🤖🧠) The Withdrawal Symptoms
When the LLM users were suddenly deprived of their AI crutch, forced to rely solely on their own minds, they underperformed cognitively with reduced alpha/beta activity and poor content recall.
This may be more than a temporary struggle: a cognitive dependence.
Their brains, accustomed to the AI doing the heavy lifting, struggled to re-engage the neural networks necessary for complex tasks.
Then the authors infer, when they wrote with ChatGPT doing most of the heavy lifting, those networks quieted down, as if the brain said, “Why the effort if the AI is doing it?”
This aligns with the concept of cognitive offloading (introduced by this and other studies), where we unconsciously relax our mental effort once we trust an external aid to handle the task.
However, some of the figures and data in the table are a lot less straightforward to interpret than the others (from someone who never touched neuroscience), hence one of the problems I will talk about later, whether there is a potential lack of samples, and also over-interpreting the results.
The Ownership and Satisfaction of Creating.
Writing also has an emotional dimension.
So the team probed how engaged students felt in their writing and how well they remembered and owned what they wrote.
You can see that the contrast between AI users and non-users has grown.
Ability to recite the previous content
After each session, students (you) were asked to quote something from their own essay and discuss their thought process. The LLM group struggled shockingly; over 80% of them could not recall a single specific phrase from the essay they had just finished, or misquoted themselves when they tried.
Not a single ChatGPT-using student could perfectly remember a line they wrote. And of course, you could have guessed, this didn’t improve by the last session. Those who had relied on AI and then had to go without it (LLM-to-Brain group) showed the same dismal recall, which the authors call “cognitive debt.”
The theory is that those who relied on AI hadn’t encoded the knowledge deeply (the AI did most of the work), and they ended up with a very shallow memory of the content. The brain-only and search groups (which you were in) fared much better! Nearly all of the non-LLM students could accurately quote or summarize their essays.
The sense of ownership of the work created
This gap (AI vs. Brain only) extended to the sense of ownership.
Many LLM-dependent students expressed little or no ownership over their work, for example
One participant claimed, "The essay was about 50% mine.
I provided ideas, and ChatGPT helped structure them.
or that
"I felt like the essay was mostly mine, except for one definition I got from ChatGPT."
We see this fragmented and conflicted sense of authorship, with many unable to discern which ideas were truly theirs. I argue they outsourced not only the writing, but lost the potential connection with their own work to AI. See the screenshot below for the ownership survey data.
There was also the “Satisfaction“ factor in play…
You would think the brand only group with display the highest satisfaction?
To be fair, there is no difference in the satisfaction factor among the three groups when we only judge this based on their “Question 6. Satisfaction with the essay.“ Even though the author claims
participants who were in the Brain-only group reported higher satisfaction…
I guess this is another reason that fueled the debates about this paper’s assertions.
Critical thinking and metacognition
An unexpected finding was that in the crossover session, the brain-to-LLM (🧠🧠🧠🤖) group (who initially only got AI access after learning to write unaided) may have actually used ChatGPT in a much more reflective way.
They suspect these students mentally compared the AI’s suggestions to their own earlier efforts, effectively double-checking or integrating ideas more thoughtfully.
The authors further support this theory by connecting the EEG readings… In plain English, those who had first honed their human skills (🧠🧠🧠🤖) seemed to use the AI as a collaborative tool, whereas those who started with AI (🤖🤖🤖🧠) used it as a crutch. The former still exercised their minds and maintained some agency; the latter just hit “auto-complete” and zoned out.
All in all, this study paints a picture of AI-induced mental laziness. The convenience of ChatGPT made students disengage from the hard parts of writing, such as remembering facts, crafting original phrases, and critically examining arguments.
Students who rely on AI may get the immediate job done, yes, but at the cost of learning less from it.
Or as the author warns
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.
AI Judge vs. Human Teachers
This isn't the first study that compared the differences between human and AI judges. But certainly something worth keeping an eye on, most people overlooked the kind of impact this could have on our daily life.
Suppose you consider an art auction.
AI might appraiser uses only technical metrics (color balance, symmetry, brushstroke consistency) while another considers emotional impact, cultural significance, and artistic innovation.
Similarly, if you think about how many companies have already started to use AI to screen CVs… AI resume screeners might favor keyword-optimized applications over truly innovative candidates.
This is so much more than 'good' or 'bad', but who decides, and how?
The dual assessment by human teachers and an AI judge revealed a fascinating, while disturbing.
They found that AI judges consistently scored essays higher than human teachers on a 0-5 scale.
While human teachers penalized formulaic writing with scores as low as 1-2 for uniqueness, the AI judge continued awarding 4s to the same pieces. This is consistent with what we have seen so far, ie, that AI tends to be a human pleaser. It even appears to mistake fluency for originality.
Which means, AI judges tend to optimize for polish and structure (as a machine should be) rather than authentic insight. They reward what looks good rather than what is good.
The Viral Research and The Critics
It’s sooooo rare for an academic preprint to go viral like this one.
I guess the experiment is easy-ish to grasp (essays with vs. without AI), and with the compelling visuals, like side-by-side brain scans. A red-hot neuron active brain vs. a cool-blue AI-addled brain is worth a thousand words.
These graphics made the “cognitive cost” of using AI tangible and shareable, helping the story spread across news sites and X.
Honestly, I’m glad this study went viral, despite its imperfections.
A messy, complicated reality about AI’s impact on cognition finally became part of the dinner table conversation. You can’t fix a problem if no one knows it exists. Viral doesn’t mean flawless, but it does mean people are finally paying attention.
The idea of cognitive offloading is NOT new. It didn’t prove anything unheard of.
It was unlucky for many previous AI and critical thinking studies that didn't get as much notice as they should have. These researchers will take hype-fueled awareness over ignorance any day.
An article I wrote early this year referenced three fascinating studies about how AI will impact our cognitive load and critical thinking.
Scientific Critique: The Usual Suspects
And of course, because it went viral, it's now under scrutiny of hundreds of thousands of people.
1. Tiny sample size
Of course, 18 people per group (and just 9 in each of the critical crossover) is too small to “prove” anything big. Welcome to the reality of every real-world cognitive neuroscience study, most labs are underfunded (though I don't know if this one is well funded), EEG is expensive, and no grad student survives wiring up 200 heads in one semester. If you’re waiting for massive n-sizes in brain studies outside pharma, we’ll all die waiting.
2. EEG Correlation ≠ Causation
EEG can reveal correlations and patterns of brain activity, but not necessarily the underlying deep causal mechanisms. The difference in neural connectivity could be due to learning, fatigue, boredom, or just different strategies. That doesn’t invalidate the trend this study and all the previous studies found: “brain rot by AI.”
3. No Peer Review
It’s a preprint. Here’s the little secret: plenty of preprints get published after peer review with only minor changes (between 70% to 90% can be marked unchanged). Unfortunately, peer review is no magic spell validating a study, not to mention that there are many researchers now using AI to peer review… kind of misses the point.
Peer review ≠ a guarantee of truth.
4. Cognitive Load Theory (CLT) Overreach
One thing that neuroscientists always struggle with is bridging the mind/brain gap. So I can see why EEG frequency bands aren’t directly mapped to CLT’s “germane” or “extraneous” load.
5. Loaded Language
The language is punchy and, at times, moralizing…
“cognitive atrophy,” “laziness”, which don’t belong in a neutral methods section. But, this is NOT unheard of … to be fair, the objective language is boring, and the very reason why people outside of the research circle rarely read any academic papers.
Time for a change!
Second Order Implications
Issues and proposals for AI using judges
When AI consistently rewards certain patterns, it creates a feedback loop.
We are in chaotic times; there are no standards or regulations for using AI. For example, when applying for a position or deciding whether to use AI in education and how to do so, remember that education is much more than just students — it also includes professors and staff.
More often than not, more and more uni staff use AI to help with their work, from annual reviews to applications.
No one knows if our writings will be read by humans or by AI.
So if there's any chance that the writing will be read by AI, would it be better if we write in a format that AI can easily parse? This is the topic I haven't discussed yet, but if you want to stand out from the answer in ChatGPT or in Perplexity, you need to play their game.
Does that mean students (or us) learn to write for the AI judge, producing increasingly uniform content that gets high scores but lacks genuine insight?
Given that AI judges reward competence over excellence. They can identify technical proficiency but struggle with the nuanced judgment that separates good from great.
This goes beyond essay grading:
Hiring algorithms that favor resume optimization over genuine qualifications
Content recommendation systems that promote engagement over accuracy or insight
Academic evaluation tools that reward gaming the system over learning
Creative platform algorithms that amplify formulaic content over original work
By understanding the differences, we can potentially design systems that complement human judgment. AI can handle scale and consistency, while humans provide context, creativity assessment, and authenticity detection.
Could there be an evaluation system that is "AI-aware", designed to account for the specific blind spots and biases that AI judges bring to the table?
Solutions to the lack of brain training: the timing of introducing AI matters.
If AI tools are introduced after a foundation of human effort, the outcomes look positive. The brain-to-LLM group’s success is a key insight. These students first grappled with problems on their own, and only then augmented their process with AI.
As a result, the students retained strong cognitive engagement and actually seemed to get the best of both worlds: their essays in Session 4 were unique and well-structured, and their brains remained highly active even with the AI assistance.
That said, we will need more experiments to see if the initial brain use is long-lasting and for how long…
Still, the research suggests a model for education (and work): do the heavy cognitive lifting yourself first, then use AI as a boost, not a crutch. Which makes sense, given what I’ve read and seen so far in how people interact with AI chatbots.
The implications also resonate with themes I explored in “The Silent Classroom.”
In that piece, I described how a classroom full of AI-dependent students becomes “silent”, as in no one asks questions or debates ideas, because everyone quietly leans on their AI helper. The “shared struggle” of learning evaporates.
The cognitive debt study shows how that silence begins in individual minds: if each student isn’t wrestling with the material (because an LLM hands them answers), they have nothing authentic to share with the class.
Over time, this could hollow out the collaborative aspect of learning. Education might shift from active engagement to a passive consumption of AI-generated answers – a troubling scenario for any teacher witnessing the lights turn off in students’ eyes.
Think about thinking like an exercise.
It is tempting to skip leg day and let a robot do the squats for you, but you’re the one who grows weaker. Or that it might be less of a hassle to train alone, but you’re the one who loses social connections.
Ultimately, the takeaway is not to banish AI tools. A hammer can build a house or bruise your toes; it all depends on whether you’ve learned to use it with proper safeguards.
But to use them wisely!
Other AI impact human behaviour studies to read:
The Widening Gap: The Benefits and Harms of Generative AI for Novice Programmers
AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking
On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial
Or please do send your favourite AI’s impact on human behaviour my way!
Share this post