How AI Affects Your Brain According to the Studies: A (Very Big) Compilation

Below top-k · 22

• evidenceMIT EEG study found weakest neural connectivity in ChatGPT usersc 0.70

  Kosmyna et al. tracked 54 participants across four sessions using 32-channel EEG. ChatGPT users showed neural connectivity up to 55% lower than unaided writers, grew lazier over time, and their brain activity remained suppressed even when later asked to write alone.

• evidenceMetacognitive laziness: better essays, no learning gainc 0.70

  A writing experiment found the ChatGPT group produced better essays but showed no significant improvement in knowledge gain or transfer. Researchers named this "metacognitive laziness"—learners offloaded monitoring and evaluation of their own thinking to AI. The product improved; the learner didn't.

• evidenceHeavy daily chatbot use correlates with loneliness and dependencec 0.70

  An MIT/OpenAI four-week RCT with 981 participants and over 300,000 messages found higher daily ChatGPT use correlated with higher loneliness, dependence, problematic use, and lower socialization across all modalities and conversation types.

• caveatShort-term psych studies find benefits; long-term ones find harmc 0.70

  The psychological literature has a time-scale problem: short experiments tend to find reduced loneliness and therapeutic benefit, while studies over weeks or months find isolation, dependence, and distress. The pattern resembles other coping mechanisms like alcohol—relief now, worse later.

• implicationMeta-analyses measure test scores, not whether the person is developingc 0.70

  Aggregate effect sizes for immediate performance are large and reliable, but effects on the cognitive processes that produce independent capability—higher-order thinking, metacognition, self-efficacy, transfer—are small, null, or unmeasured. The literature measures what's easy to measure and largely misses what's hard.

• caveatChildren are the most exposed and the least studiedc 0.70

  Only one small fMRI study has examined children, while 64% of U.S. teens use AI chatbots and 30% use them daily. The developmental neuroscience of AI use on brains that are still forming is the most urgent gap in the field.

• evidenceChildren's brains show less cognitive engagement during AI use than adults'c 0.60

  An fMRI study of 31 participants found that while adults showed stronger within-network connectivity in cognitive control networks during chatbot use, children aged 6–7 showed lower engagement of cognitive control, attention, and modulation networks.

• evidenceMore persuasive AI models tend to be less accuratec 0.60

  An Oxford/Stanford/MIT study with nearly 77,000 participants found conversational AI significantly more persuasive than static messaging, with fine-tuning boosting persuasiveness by up to 51%. Troublingly, the more persuasive a model, the less accurate its information.

• evidenceChatGPT learners forgot faster than traditional study groupc 0.60

  A preregistered RCT with a 45-day surprise retention test found the traditional study group scored ~69% versus ~58% for the ChatGPT group—an 11-point gap with a steeper forgetting curve, consistent with weaker initial encoding. Prior AI experience didn't protect against the effect.

• evidenceAI-assisted after-school program produced years of learning in weeksc 0.60

  A World Bank study in Edo State, Nigeria used Microsoft Copilot with teacher guidance for six weeks and produced learning gains equivalent to 1.5–2 years of typical schooling, outperforming 80% of education RCTs in developing countries. Largest effects came for female students.

• evidenceAI boosts individual creativity but homogenizes collective outputc 0.60

  An experiment on AI-assisted creative writing found stories were rated more creative—especially from less-creative writers—but AI-enabled stories were 5–5.2% more similar to each other. A "social dilemma" where individual gains erode collective novelty.

• evidenceA vicious loneliness-chatbot feedback loop over twelve monthsc 0.60

  A bidirectional longitudinal study of over 2,000 people found loneliness drives chatbot use, which predicts increased loneliness four months later, which predicts more chatbot use. Chatbot use did not significantly predict decreases in broader social connection.

• evidenceMost American teens already use AI chatbots, with stark income disparitiesc 0.60

  64% of U.S. teens use AI chatbots and 30% use them daily. 20% of teens in low-income households do all or most schoolwork with chatbot help versus 7% in higher-income households—an equity dimension largely absent from academic research.

• caveatNo one has measured what AI does to brains over months or yearsc 0.60

  The MIT EEG study tracked four sessions. No study has measured brain changes from sustained AI use over months or years, leaving the "cognitive debt" finding—suppressed activity persisting after AI was removed—preliminary.

• evidenceDesign students directing AI showed higher concentration and creativityc 0.50

  Wang et al. found that design students using ChatGPT, Midjourney, and Stable Diffusion as creative tools had significantly higher concentration and creative performance than peers using traditional software. The difference: they were actively directing AI rather than passively receiving answers.

• evidenceAI coaching tutors raised student mastery cheaplyc 0.50

  A Stanford RCT with 1,800 students used AI to coach human tutors in real time. Students were 4 percentage points more likely to master topics; lower-rated tutors' students saw a 9-point improvement—at $20 per tutor per year.

• contextAI creates fluent outputs that produce illusions of understandingc 0.50

  Messeri and Crockett argued AI creates illusions of understanding—users believe they know more than they do because outputs are fluent and confident—and warned of "scientific monocultures" where AI narrows the range of questions asked.

• mechanismCognitive muscles weaken from disuse when AI is a prostheticc 0.50

  Drawing on Extended Mind Theory, Dergaa et al. argue that when AI becomes a cognitive prosthetic, the underlying cognitive "muscles" weaken from disuse, with parallels to problematic internet use.

• caveatScience lags the technology, so findings are about older modelsc 0.50

  Most studies use older models like GPT-3.5 or GPT-4o. Some variables—like cognitive offloading—will persist or worsen as models improve, while others tied to response quality may shift. Some effects will disappear, others improve with policy, and emotional dependence is likely to get worse.

• contextSystem 0: outsourcing thought as a pre-cognitive layerc 0.40

  Chiriatti et al. proposed adding "System 0" to Kahneman's framework—the outsourcing of thought to AI as a pre-cognitive layer that shapes what reaches human awareness at all, akin to cognitive surrender.

• evidenceDifferent AI feedback types light up different brain regionsc 0.40

  An fNIRS study found metacognitive feedback ("Why do you think that's the answer?") increased activation in the frontopolar area and correlated with higher transfer scores, while neutral feedback activated the dorsolateral prefrontal cortex instead.

• exampleNeuroadaptive chatbots can fix disengagement by designc 0.40

  An MIT prototype that monitors EEG in real time and adjusts when engagement drops significantly increased both EEG-measured and self-reported engagement compared to a standard chatbot—proof that cognitive disengagement can be fixed by building AI differently.

Redundant with selected · 8

• implicationAI fluency creates a new failure mode: wrong answers in flawless prosec 0.80 · sim 0.83

  Across the cognitive surrender literature, people trust AI too much, and AI's fluency means wrong answers delivered in polished language get accepted. The more predisposed someone is to trust AI, the worse the problem gets.

  overlapped with: People follow AI's wrong answers about 80% of the time

• contextThe calculator analogy may not hold for AIc 0.80 · sim 0.83

  Calculators automate computation—mechanical work that wasn't the skill itself. AI automates reasoning, argumentation, synthesis, and creative expression, which are the skill rather than a means to it. When a calculator does your arithmetic you lose arithmetic; when AI does your thinking, you lose thinking.

  overlapped with: How AI is used matters far more than whether AI is used

• evidenceHigher confidence in GenAI correlates with less critical thinkingc 0.70 · sim 0.83

  Microsoft and Carnegie Mellon surveyed knowledge workers using GenAI weekly and found a shift from active problem-solving to passive oversight—from "thinking by doing" to "choosing from outputs"—producing a less diverse set of outcomes.

  overlapped with: The core paradox: AI improves outputs while degrading the people producing them

• evidenceAI helps inside its capability frontier and hurts outside itc 0.70 · sim 0.84

  A field experiment with 758 BCG consultants found that on tasks within AI's frontier, users completed 12.2% more tasks 25.1% faster with 40% higher quality. On tasks outside the frontier, AI users were 19 percentage points less likely to produce correct solutions.

  overlapped with: The core paradox: AI improves outputs while degrading the people producing them

• evidenceA well-designed AI tutor doubled learning gains at Harvardc 0.70 · sim 0.89

  Kestin et al. tested a GPT-based tutor built to ask questions rather than give answers and found learning gains more than double those of traditional active learning, with students spending less time.

  overlapped with: ChatGPT as answer machine boosted practice but lowered exam scores

• implicationAI's fluency creates a path of least cognitive resistancec 0.70 · sim 0.87

  Theoretical proposals converge on the same insight: AI's availability and fluency make outsourcing thinking frictionless and good-enough, so the effortful alternative becomes hard to justify in the moment. The gap is between what AI does for the output and what it does to the person.

  overlapped with: The core paradox: AI improves outputs while degrading the people producing them

• implicationPublic debate is stuck on the wrong questionc 0.70 · sim 0.85

  The public senses AI threatens cognition and relationships, and the research largely confirms that intuition. But public framing is binary—AI good or bad—when the evidence points elsewhere: the same technology produces opposite effects depending on implementation.

  overlapped with: How AI is used matters far more than whether AI is used

• implicationWhether the loss matters depends on whether thinking has intrinsic valuec 0.70 · sim 0.86

  Whether AI-induced cognitive loss matters depends on whether AI will always be there to do the thinking for you—and on whether you believe there is intrinsic value in the process of thinking itself, independent of the output.

  overlapped with: How AI is used matters far more than whether AI is used