Your Brain on AI

Today I was doing what millions of us do without thinking much about it: letting Instagram and YouTube Shorts decide what I would see next. Up came a short video titled The Machine That Thinks For You: Your Brain on ChatGPT — Professor Jason Kohler. The clip is here: https://youtube.com/shorts/zAzjOJi18A0. It was only a couple of minutes long, but its message was depressing enough to stop me.

The video said MIT researchers had put 32-electrode EEG caps on people writing essays. One group used ChatGPT, another used a search engine, and a third worked unaided. According to the video, the unaided writers showed the richest brain connectivity and the ChatGPT users the weakest. It moved from there through the “Google effect,” GPS and London taxi drivers, and finally to doctors whose unaided detection performance apparently deteriorated after working with an AI system. The closing implication was hard to miss: every time a machine gives us an answer, perhaps we pay for it with a little bit of our own cognitive development.

That bothered me, partly because it does not match my own experience of using generative AI. Sometimes it is easy. Sometimes it is absurdly difficult. I have spent long sessions chasing a problem through hallucination after hallucination, rephrasing a question, checking primary sources, rejecting an answer, asking another question prompted by the previous answer, and gradually discovering that the real question was not quite the one I started with. That does not feel like going mentally to sleep. Quite often it feels like intellectual exercise.

The RAG-check clerk

That question sent my mind somewhere unexpected: Laurie Anderson's Let X=X, from her 1982 album Big Science. I remembered the wonderfully deadpan image of a hat-check clerk and started playing with the idea of a different sort of clerk for the age of Big AI. Not a quotation of Anderson's lyrics, but a little response to the memory of them.

The joke turned out to contain the method. The machine can retrieve, compare, summarize and challenge, but somebody still has to take responsibility for deciding whether the claim matches the ticket. At the end of the line, the RAG-check clerk is me.

What I mean by a RAG check

First: the Short is not nonsense

The most important finding from checking the video is that its factual backbone is surprisingly strong. The MIT study exists. It did compare LLM, search-engine and brain-only essay-writing conditions. Fifty-four people participated in the first three sessions, while only eighteen completed the fourth crossover session. The researchers reported the strongest and most distributed EEG connectivity in the brain-only condition, intermediate connectivity in search users and the weakest connectivity in the LLM group. They also reported reduced alpha and beta connectivity when earlier LLM users later wrote without the tool.

But one linguistic shortcut matters enormously: EEG connectivity is not a generic “brain activity meter.” The study measured relationships among electrical signals across frequency bands and scalp electrodes during a particular writing task. Calling one condition “less active” is understandable shorthand, but it invites a stronger conclusion than the measurement alone supports. Lower connectivity during one task is not the same as brain damage, lost intelligence or general cognitive deterioration.

There is another reason for caution. The MIT work was released as a preprint and has attracted methodological criticism. A 2025/2026 commentary by Stanković and colleagues praised the importance of the research while raising concerns about sample size, reproducibility, aspects of the EEG analysis, reporting consistency and transparency. That does not invalidate the MIT results. It does mean they should be treated as provocative early evidence rather than the final word on what ChatGPT does to the human brain.

The older story: cognitive offloading is real

The Short's 2011 memory reference is also real. Sparrow, Liu and Wegner's famous Science paper examined what became popularly known as the “Google effect.” When people expected information to remain available externally, memory shifted: they were less likely to remember the information itself and better able to remember where it could be found. That belongs to a larger literature on transactive memory and cognitive offloading.

The sensational version would be: “Google made us stop remembering.” The more interesting version is: human beings change what they choose to remember when reliable external memory exists. That can be a loss, but it can also be an adaptation. I no longer remember dozens of telephone numbers because my phone does. That freed memory is not necessarily destroyed; the question is what I do with the cognitive capacity I no longer spend on telephone numbers.

Navigation gives us a similar warning. Work by Javadi and colleagues showed hippocampal and prefrontal recruitment during active navigation and route planning. Dahmani and Bohbot later reported that greater habitual GPS use was associated with poorer spatial memory during unaided navigation. And the London taxi-driver literature is among the classic demonstrations of adult neuroplasticity: experienced taxi drivers, and later people learning “the Knowledge,” showed structural hippocampal differences associated with demanding spatial learning.

Again, the safe conclusion is not “GPS shrinks your brain.” It is that abilities we repeatedly exercise can be strengthened, while abilities routinely delegated to external systems may receive less practice. That is a much more defensible principle, and it transfers naturally to AI.

The doctors: probably the most unsettling claim

The colonoscopy claim sounded almost too neat, but it checks out. Budzyń and colleagues examined endoscopists working in centres that introduced AI-assisted colonoscopy. When the doctors subsequently performed standard colonoscopy without AI, adenoma detection was reported at 22.4% after AI exposure compared with 28.4% before it—a fall of six percentage points.

That deserves attention, but the word observational matters. This was not proof that AI exposure directly damaged the doctors' diagnostic ability. Other temporal, procedural and behavioural factors can contribute. The researchers themselves used cautious causal language.

And here the complete literature creates a fascinating paradox: meta-analyses of AI-assisted colonoscopy find that AI can increase adenoma detection while it is being used. So the machine may improve the human-machine system at the same time that reliance on it creates a risk that unaided human performance weakens. That is not a contradiction. It is exactly the sort of trade-off aviation, medicine, navigation and industrial automation have confronted for decades.

Assisted performance and human learning are not the same outcome. A machine can make the team better today without necessarily making the human better tomorrow.

Now turn the telescope around

If we stop the investigation there, the YouTube Short wins by default. But that would leave out a rapidly growing body of evidence showing that AI can also be deliberately designed as a cognitive scaffold.

Creativity: more creative individually, perhaps more ordinary collectively

Doshi and Hauser's 2024 Science Advances experiment is one of the clearest demonstrations of the double-edged effect. Writers who had access to generative-AI ideas produced short stories that independent evaluators judged, on average, to be more creative, better written and more enjoyable. Less-creative writers benefited particularly strongly.

Yet the AI-assisted stories also became more similar to one another. Later work on brainstorming has reinforced concern about reduced idea diversity. In other words, AI can raise the floor of individual creativity while narrowing the collective distribution of ideas. That is an extraordinary result because both optimists and pessimists can quote it accurately.

It also explains much of what people complain about when they describe AI-generated infographics, essays or posts as having a recognizable sameness. The individual object can be polished while the cultural ecosystem becomes more homogeneous. The solution is not necessarily to abandon AI. It may be to make the human originate more of the conceptual direction before asking the machine to expand it.

Learning: when the AI is designed to teach rather than answer

A 2025 randomized controlled trial by Kestin and colleagues in an undergraduate physics course provides a particularly important counterexample to the “AI makes us intellectually lazy” narrative. Students using a carefully engineered AI tutor learned significantly more than students receiving an in-class active-learning lesson. The AI students also spent less median time on the lesson and reported greater engagement and motivation.

Figure 1 puts the central learning result into numerical form: the reported median post-test score was 4.5 in the AI-tutor condition versus 3.5 in the active-learning classroom, from a combined pre-test median of 2.75.

The word carefully matters. This was not simply a blank chatbot window. The system was designed around established pedagogical principles: active engagement, cognitive-load management, scaffolding, targeted feedback, accuracy controls and self-pacing. The researchers explicitly warned that hallucinations were a problem and supplied structured solutions and sequencing to keep the tutor on track.

Figure 2 shows that the larger learning gain did not come from simply keeping students at the computer longer: median time on task was 49 minutes for the AI tutor versus the 60-minute class period.

That is almost the mirror image of the lazy-use scenario. Instead of “give me the answer,” the machine participates in the same sort of feedback loop that makes one-to-one tutoring powerful.

Figure 3 adds the students' subjective response: mean engagement and motivation ratings were both higher in the AI-tutor condition, with statistically significant differences reported by the authors.

AI can even make the human teacher better

Stanford's Tutor CoPilot pushes the argument another step. More than 700 tutors and about 1,000 students participated in the reported randomized trial. Students whose tutors had AI assistance were about four percentage points more likely to master the mathematics topic; among lower-rated tutors, the gain was about nine percentage points. Analysis of hundreds of thousands of tutoring messages found that AI-assisted tutors were more likely to use probing questions and less likely to rely on generic praise.

Figure 4 shows the overall and subgroup effects side by side. The larger result for lower-rated tutors is especially interesting because it suggests that AI support may sometimes narrow differences in baseline human expertise.

This is not merely an AI giving students answers. The AI appears to alter the human tutor's behaviour in a direction associated with better pedagogy. In effect, expertise is being scaffolded in real time. That is a much stronger example of augmentation.

But human + AI is not magic

A major 2024 Nature Human Behaviour systematic review and meta-analysis by Vaccaro, Almaatouq and Malone examined when humans and AI actually combine well. The important lesson is not that “human plus AI always wins.” Complementarity depends on the task, the relative abilities of human and machine, and how responsibility is divided. Sometimes the combination beats humans alone without beating the better of human or AI operating independently.

Figure 5 makes that distinction visible. Across the meta-analysis, human–AI systems substantially outperformed humans alone on average, yet failed to outperform the better standalone performer on average.

Figure 6 then separates task types. Creation tasks leaned positive for human–AI synergy, while decision tasks leaned negative—one reason a single headline about “AI performance” is misleading.

Figure 7 shows perhaps the most important moderator: when the human was stronger than the AI operating alone, combining them tended to help; when the AI was already stronger than the human, adding the human tended to reduce performance relative to the stronger standalone AI.

That finding is useful because it protects us from replacing one simplistic slogan with another. “AI makes you smarter” is no better scientifically than “AI makes you stupid.”

Performance is not intellect

This may be the central distinction in the entire discussion. If I ask an AI to write a flawless report and submit it untouched, my performance may look better. Nothing in that result proves that my knowledge, reasoning or writing skill improved.

Evidence of genuine learning becomes much stronger when the human can later perform without the AI: remember the concept, explain it in a different way, solve a novel problem, detect an error, transfer the principle to another field or reconstruct the argument after a delay. Those are the outcomes future AI-cognition studies need to measure much more often.

The same caution works in reverse. Less measured neural effort is not automatically intellectual decline. Expertise often makes difficult tasks more efficient. A chess master does not necessarily need to consciously search every branch a novice searches. So a neurological result must be interpreted alongside retention, transfer, accuracy, strategy and behaviour.

Why my own experience feels different

There are certainly times when generative AI encourages laziness. We have all seen text posted by people who apparently did not even read what the system generated. Hallucinated citations remain in place. Strange phrases remain in place. Sometimes the user has outsourced not just writing but judgment. The very low friction of generation can also create a loop of prompt, reward, post, repeat that starts to resemble the compulsive dynamics of social media.

But there is another mode. The AI produces something wrong. I notice it. I ask why. It produces a plausible explanation that still does not fit. I find the original paper. A term in that paper raises another question. That sends me to a second source which contradicts the first interpretation. I reformulate the question and force the machine to distinguish what is known from what is inferred. Eventually the output may be only a small fraction of the work that happened in the conversation.

That mode can also be addictive in the ordinary sense that an absorbing intellectual activity is hard to put down. But subjectively it resembles falling down a library rabbit hole more than scrolling a feed. It creates questions faster than it closes them.

The important caveat is that my experience is anecdotal. I cannot infer from feeling mentally stimulated that my long-term cognition is improving. But the tutoring and creativity research establishes that it is at least scientifically plausible for AI-mediated interaction to increase learning or creative performance when the human remains cognitively engaged.

A practical cognitive workout for AI

Based on the evidence reviewed here, I would use seven working rules—not commandments, and not claims that each has been separately validated as a complete protocol.

1Think before prompting. Generate your own explanation, hypothesis, outline or ideas first when the purpose is learning or creativity.

2Ask for resistance. Tell the AI to challenge assumptions, generate counterarguments, expose missing evidence and point out where an argument can fail.

3Use clues before answers. For skills you want to retain, ask the system to scaffold the problem one step at a time rather than solve it immediately.

4Explain it back. After learning something with AI, reconstruct it in your own words without copying. Better still, teach it.

5RAG consequential claims. Retrieve the original source, distinguish it from commentary, and make yourself decide whether the evidence actually supports the sentence.

6Periodically remove the tool. Try the same category of task unaided. If performance collapses, the tool may be concealing rather than building competence.

7Stay the final judge. AI may be researcher, critic, tutor, simulator, editor or adversary. It should not quietly become the person responsible for deciding what you believe.

Matched counter-RAG: ten claims for cognitive augmentation

The original Short presents a sequence of claims pointing toward cognitive offloading, reduced engagement and possible deskilling. To keep the comparison symmetrical, the next ten atomic claims are drawn from the strongest counter-evidence discussed above. They are not slogans such as “AI makes you smarter.” Each is deliberately narrow enough to be tested against the underlying study.

What the ten counter-claims actually establish

Taken together, they provide strong evidence for a narrower proposition than “AI boosts intellect.” Generative AI can, under particular interaction designs, improve measured creative output, increase learning, increase engagement and motivation, improve the behaviour of human tutors, and augment human task performance. The strongest educational examples are deliberately structured so that the AI scaffolds the learner rather than simply replacing the learner's reasoning.

Just as importantly, the counter-evidence contains its own warnings. Doshi and Hauser found reduced collective diversity even while individual stories improved. Vaccaro, Almaatouq and Malone found substantial human augmentation compared with humans alone, but no average synergy against the better of human or AI alone. In other words, the positive literature itself argues against blind enthusiasm: interaction design, task type, baseline human ability and the division of labour matter.

That gives us two matched RAG ledgers: evidence that AI can remove cognitive work, and evidence that AI can increase the quality of human learning and creation. The contradiction largely disappears when we stop treating every form of “AI use” as the same intervention.

Atomic RAG ledger

These scores summarize the evidence examined for this article. They are deliberately conservative. A high score means “strong support for this precise formulation,” not “the topic is scientifically settled forever.”

So, am I damaging myself?

The scientifically responsible answer is: we do not know enough to answer that at the level of an individual over years of use. Anyone who tells me with confidence that routine generative-AI use will either increase my intellect or rot my brain is running ahead of the evidence.

What we can say is more useful. Cognitive offloading is real. Skills that are not exercised can deteriorate. Automation can create deskilling and automation bias. Generative AI can homogenize creative output. The MIT essay study raises a legitimate warning about passive reliance.

But we can also say that carefully structured AI tutoring can increase measured learning, AI assistance can improve human teaching behaviour, generative AI can raise individual creative performance, and human-AI systems can outperform human-only work in appropriate settings. The evidence therefore points away from the machine as the single causal variable and toward the interaction.

The healthiest AI may not be the machine that gives me the best answers. It may be the machine that makes me ask better questions—and leaves me capable of answering them when the machine is gone.

What we still do not know

We do not have twenty-year longitudinal studies of generative-AI use because the technology has not existed in its present form for twenty years. We do not know whether intensive augmentation-style use produces durable changes in broad reasoning ability. We do not know the optimal balance between offloading routine cognition and deliberately exercising it. We do not know whether effects differ systematically with age, expertise, neurodiversity, educational background or task domain. And we are still learning how social incentives—speed, grades, likes, workplace quotas and the pressure to publish—change the way people choose to use these systems.

Those unanswered questions are not weaknesses in the article. They are part of the result. A proper RAG check does not turn uncertainty into a confident conclusion merely because confidence makes a better meme.

Glossary

Automation bias

The tendency to over-trust automated recommendations or fail to notice errors because an automated system appears authoritative.

Cognitive augmentation

Use of a tool that expands, scaffolds or intensifies human reasoning rather than simply replacing it. Used here as an analytical framework, not a single standardized clinical construct.

Cognitive debt

The phrase used by Kosmyna et al. to describe possible downstream cognitive costs associated with reliance on an AI assistant in their essay-writing paradigm.

Cognitive offloading

Using an external aid—paper, calculator, search engine, GPS, phone or AI—to reduce internal cognitive demands.

Cognitive substitution

The term used in this article for cases where the AI performs a cognitive operation the human would otherwise have performed.

EEG connectivity

Statistical relationships among electrical brain signals measured at different scalp electrodes and frequency bands. It should not be casually equated with an overall quantity of “brain activity.”

Generative AI

AI systems capable of producing new text, images, audio, code or other content based on learned patterns and user inputs.

Hallucination

A fluent AI output that contains unsupported, fabricated or incorrect information despite being presented plausibly.

RAG

Retrieval-augmented generation: grounding generated responses in retrieved external material. In this essay, “RAG check” additionally refers to atomic claim decomposition and source verification.

Socratic scaffolding

A teaching approach that uses questions, hints and structured prompts to make the learner perform the reasoning rather than simply receive an answer.

Transactive memory

A system in which people remember not only information itself but also where or from whom that information can be retrieved.

Transfer

The ability to apply knowledge or skill learned in one setting to a new problem or context—an important distinction from merely performing well with assistance.

Bibliography and source map

The list deliberately includes sources pointing in both directions: cognitive offloading and deskilling, as well as learning, creativity and augmentation. Where possible, the original study rather than a news summary is linked.

1. Kosmyna et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab / arXiv preprint. https://arxiv.org/abs/2506.08872
2. MIT Media Lab (2025). Your Brain on ChatGPT project/publication page. MIT Media Lab. https://www.media.mit.edu/publications/your-brain-on-chatgpt/
3. Stanković et al. (2025/2026). Comment on: Your Brain on ChatGPT. arXiv methodological commentary. https://arxiv.org/abs/2601.00856
4. Sparrow, Liu & Wegner (2011). Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips. Science 333(6043). https://doi.org/10.1126/science.1207745
5. Javadi et al. (2017). Hippocampal and prefrontal processing of network topology to simulate the future. Nature Communications 8. https://doi.org/10.1038/ncomms14652
6. Dahmani & Bohbot (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports 10. https://doi.org/10.1038/s41598-020-62877-0
7. Maguire et al. (2000). Navigation-related structural change in the hippocampi of taxi drivers. PNAS 97(8). https://pubmed.ncbi.nlm.nih.gov/10716738/
8. Woollett & Maguire (2011). Acquiring 'the Knowledge' of London's layout drives structural brain changes. Current Biology. https://pubmed.ncbi.nlm.nih.gov/22169537/
9. Budzyń et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy. The Lancet Gastroenterology & Hepatology. https://pubmed.ncbi.nlm.nih.gov/40816301/
10. Makar et al. (2025). Use of artificial intelligence improves colonoscopy adenoma detection. Gastrointestinal Endoscopy meta-analysis. https://pubmed.ncbi.nlm.nih.gov/39216648/
11. Soleymanjahi et al. (2024). Artificial Intelligence-Assisted Colonoscopy for Polyp Detection: A Systematic Review and Meta-analysis. Annals of Internal Medicine. https://pubmed.ncbi.nlm.nih.gov/39531400/
12. Doshi & Hauser (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances 10(28). https://doi.org/10.1126/sciadv.adn5290
13. Kestin et al. (2025). AI tutoring outperforms in-class active learning: an RCT. Scientific Reports. https://doi.org/10.1038/s41598-025-97652-6
14. Wang et al. / Tutor CoPilot (2024). Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise. Stanford National Student Support Accelerator. https://nssa.stanford.edu/studies/tutor-copilot-human-ai-approach-scaling-real-time-expertise
15. Vaccaro, Almaatouq & Malone (2024). When combinations of humans and AI are useful: a systematic review and meta-analysis. Nature Human Behaviour 8. https://doi.org/10.1038/s41562-024-02024-1
16. Meincke, Nave & Terwiesch (2025). ChatGPT decreases idea diversity in brainstorming. Nature Human Behaviour 9. https://doi.org/10.1038/s41562-025-02173-x
17. Yilmaz et al. (2024). Real-Time multifaceted artificial intelligence vs In-Person instruction in surgical simulation. Scientific Reports. https://doi.org/10.1038/s41598-024-65716-8
18. Létourneau et al. (2025). A systematic review of AI-driven intelligent tutoring systems in K-12 education. npj Science of Learning. https://doi.org/10.1038/s41539-025-00320-7
19. Wu et al. (2025). Human-generative AI collaboration enhances task performance but undermines intrinsic motivation. Scientific Reports. https://doi.org/10.1038/s41598-025-98385-2
20. Nonesuch Records. Laurie Anderson — Big Science / Let X=X. Artist/label reference. https://www.nonesuch.com/albums/big-science-lp

Copyright © Robert Harvey Curtis (rhc)

Powered by the Solynth Network

https://public.solynth.net