Practical AI · research brief · all sources verified against primary documents June 18, 2026
What AI Is Actually Doing to Your Brain
The evidence, by user type — cognition AND wellbeing — separating real findings from the headlines.
The bottom line
The strongest finding is not "AI makes you dumber." It's that how you use it is the variable. Passive, copy-paste, "trust it blindly" use shows reduced mental engagement and weaker recall of your own work. Active, verify-and-build-on-it use looks neutral to beneficial.
Solid: AI offloads thinking and improves short-term output; in learning, easier-now can mean weaker-without-it-later. Thin: permanent brain harm in healthy adults, and almost everything about kids and the long term. The viral "ChatGPT rots your brain" claims trace to one small preprint that did not show what the headlines said.
Peer-reviewed
Preprint — not peer-reviewed
Survey / self-report
Review / synthesis
Retracted
1. Passive / light everyday users
Cognitive offloading, over-reliance, the "Google effect" applied to chatbots.
Sparrow, Liu & Wegner — "Google effects on memory"Peer-reviewed
Columbia / Harvard · 2011 · Science 333(6043) · 4 lab experiments (behavioral, not brain imaging)
The foundation. When people expect information to stay accessible, they remember less of the content and more of where to find it. This is the offloading mechanism the whole AI debate now extends.
Note: it's about search engines, not chatbots. Applying it to AI is reasonable but not yet directly proven for everyday chatbot use.
science.org/10.1126/science.1207745
Chirayath, Premamalini & Joseph — "Cognitive offloading or cognitive overload?"Review
Frontiers in Psychology · 2025 · PMC12678390 · narrative review
AI has dual potential: it can free mental resources or erode autonomy and engagement when it decides too much for you. Effects depend on design and how much agency the user keeps.
pmc.ncbi.nlm.nih.gov/PMC12678390
2. Power users / knowledge workers & professionals
Deskilling vs augmenting. The clearest pattern: heavy reliance shifts you from generating to verifying.
Lee et al. — GenAI and critical thinking at workPeer-reviewed Self-report
Microsoft Research + Carnegie Mellon · 2025 · ACM CHI · n=319 knowledge workers, 936 real examples
Higher confidence in the AI predicted less critical-thinking effort. GenAI lowers perceived effort and shifts the work from gathering/problem-solving toward verification and oversight.
Self-reported and correlational — it measures what people say they do, not a tested outcome.
microsoft.com/research
Gerlich — AI tool use and critical thinkingPeer-reviewed Survey
SBS Swiss Business School · 2025 · Societies 15(1):6 (MDPI) · n=666 · survey + mediation analysis
Frequent AI use was negatively correlated with critical-thinking scores, with cognitive offloading as the mediator. Younger participants relied more.
Correlational and self-reported — it shows an association, not that AI causes the decline.
mdpi.com/2075-4698/15/1/6
Khan & Suhluli — GenAI immersion, cognitive load & fatiguePeer-reviewed Survey/SEM
2025 · Technologies 13(11):486 · n=998 researchers · structural equation modeling
The closest evidence to "more AI = more fried." Deep GenAI immersion amplified the negative impact of cognitive load and task fatigue on work quality — heavy use added strain rather than removing it.
Self-reported and correlational, but a large sample and the most direct study of heavy-user fatigue.
mdpi.com/2227-7080/13/11/486
"AI brain fry" — HBR / BCG worker surveySurvey / report
Harvard Business Review + BCG-affiliated · 2026 · ~1,488 US full-time workers
Coined "AI brain fry": cognitive/decision fatigue from constantly overseeing and checking AI output. Names the heavy-user crash as a real, widespread workplace pattern.
Practitioner survey, NOT peer-reviewed. Real signal that the experience is shared; not a controlled or brain-measured study.
hbr.org/2026/03/when-using-ai-leads-to-brain-fry
3. Students, children & adolescents
The most important group for the long term — and where the evidence is most fragile. The pattern: better on the task with AI, sometimes worse without it later.
Kosmyna et al. — "Your Brain on ChatGPT" (the famous EEG study)Preprint
MIT Media Lab · 2025 · arXiv:2506.08872 · n=54 (18 per group; only 18 did session 4) · 32-channel EEG, essay writing over ~4 months
What it showed: brain connectivity scaled with how much help you had — "brain-only" writers had the widest networks, search was middle, the ChatGPT group the weakest. The AI group also recalled and "owned" their own essays less, and wrote more alike.
What it did NOT show: no "47% brain collapse" (that's media invention, not in the paper), no brain damage, no permanent decline. It's a small, preprint, single-task (timed essays), young/educated sample — short-term, not long-term. The lead author has publicly pushed back on the alarmist framing.
arxiv.org/abs/2506.08872
Suriano et al. — ChatGPT can promote critical thinkingPeer-reviewed
2025 · Learning and Instruction 95:102011 · n=213 Italian university students
The positive counterpoint. With active engagement, trust, and reflection, ChatGPT use was linked to better complex critical thinking. Engagement mattered more than prior knowledge. Active use ≠ passive use.
Shows the upside is real when AI is a thinking partner, not a substitute.
doi.org/10.1016/j.learninstruc.2024.102011
Zafar — ChatGPT and secondary studentsSurvey
2025 · ~300 secondary students · self-report, r=0.52
Frequent users perceived a gain in critical thinking.
Weak: self-reported perception only, and a low-tier publication venue. Treat as a weak signal, not evidence.
Wang & Fan — meta-analysis of ChatGPT & learningRETRACTED
Humanities and Social Sciences Communications · published 2025, retracted April 2026
Claimed broad learning/higher-order-thinking benefits across 51 studies. Retracted over flaws that undermined the analysis. Do not cite it as evidence — included here only so it isn't mistakenly reused.
On children & developing brains specifically: there is almost no direct, high-quality evidence on generative-AI chatbots. Concerns are extrapolated from the offloading literature and general digital-tool research. This is the biggest gap in the field.
4. Other groups — coders, older adults
Where it gets task-specific: the same tool can hurt skill-building and help productivity, depending on who and what.
Shen & Tamkin (Anthropic) — AI assistance & coding-skill formationPreprint / company research
Anthropic · 2026 · arXiv:2601.20245 · RCT · n=52 (mostly junior) developers learning an unfamiliar Python library
A real randomized trial. Developers who used AID to learn a new library scored ~17% lower on a comprehension quiz (50% vs 67%, d≈0.74), biggest gap on debugging. Productivity gains weren't statistically significant. AI can hinder forming a brand-new skill.
Small sample, recent, company research (not yet journal peer-reviewed). It's about learning, not everyday expert work.
anthropic.com/research
METR — AI's impact on experienced developer productivityPreprint
METR · 2025 · arXiv:2507.09089 · RCT, experienced open-source developers
The one people conflate with the Anthropic study. It's different: early-2025 AI tools slowed experienced developers ~19% on real tasks — even though they expected a speedup. About productivity, not learning.
Tool- and time-specific (early 2025); METR notes newer tools show smaller/mixed effects.
arxiv.org/abs/2507.09089
Pergantis et al. — chatbots & executive functionSystematic review
2025 · Brain Sciences 15(1):47 (MDPI) · systematic review across age groups incl. older adults
The assistive side. Structured, goal-directed chatbot interactions can support executive function (memory training, cognitive support), including in older adults — when the tool is designed for it, not passive everyday chat.
Evidence that intentional, structured AI use can help, not just harm.
mdpi.com/2076-3425/15/1/47
5. Wellbeing & mental health
A separate question from cognition — and the evidence here is more developed. The split: structured therapy bots help; open-ended companion bots are mixed; heavy emotional reliance is the risk.
Heinz, Jacobson et al. — "Therabot" therapy-chatbot RCTPeer-reviewed
Dartmouth (Geisel) · 2025 · NEJM AI · RCT · n=210 (106 Therabot vs 104 waitlist) · pre-registered NCT06013137
The strongest single finding in this whole topic, and it's positive. A generative-AI therapy bot produced large symptom reductions — depression d≈0.85, anxiety d≈0.79, eating-disorder risk d≈0.63–0.82 — and users rated the therapeutic bond comparable to a human therapist. First RCT of a fully generative AI therapy chatbot.
Gold-standard method (RCT, peer-reviewed). The caveat: it's a purpose-built, guard-railed clinical tool — not ChatGPT, and not a companion app.
ai.nejm.org/10.1056/AIoa2400802
Feng et al. — meta-analysis of AI conversational agents for mental healthMeta-analysis
2025 · JMIR · 15 trials, n≈1,974
AI chat agents had a moderate-to-large effect on depression (g≈0.61), strongest in subclinical groups. But effects on anxiety, stress, and general wellbeing were not significant after bias adjustment.
Tempers the hype: the depression signal is real; the broader "AI improves wellbeing" claim is not supported.
jmir.org/2025/1/e69639
Maples et al. — Replika, loneliness & suicide mitigationPeer-reviewed Self-report
Stanford · 2024 · npj Mental Health Research · n=1,006 student Replika users
The mixed picture for companion apps. Users were lonelier than typical students yet reported high perceived social support; 3% (30 people) said Replika halted their suicidal ideation.
Self-reported and a self-selected user base — the suicide-mitigation number is striking but not a controlled outcome. Benefit signal and dependence risk live side by side.
nature.com/s44184-023-00047-6
Phang, Kosmyna et al. — "affective use" & emotional wellbeing on ChatGPTPreprint
OpenAI + MIT Media Lab · 2025 · arXiv:2504.03888 · RCT ~1,000 over 28 days + 3M+ conversations + 4,000-user survey
The risk side. Very heavy use correlates with more self-reported emotional dependence, and a small group of "power affective users" account for most of the emotional interaction. Voice-mode effects depend on the user's starting emotional state and how much they use it.
Preprint (not peer-reviewed); correlational on the platform-data side. Run by OpenAI itself — read with that in mind.
arxiv.org/abs/2504.03888
Duong et al. — compulsive ChatGPT use, anxiety, burnout, sleepPeer-reviewed Survey
2024 · Acta Psychologica 251:104622 · survey + serial mediation
The dependence edge, measured. Compulsive ChatGPT use linked to anxiety → burnout → sleep disturbance. Heavy/compulsive patterns track with worse mental health.
Self-reported and correlational — shows association, not that AI causes it.
pubmed.ncbi.nlm.nih.gov/39647449
Maral et al. — "Problematic ChatGPT Use Scale"Peer-reviewed
2025 · Int. Journal of Mental Health and Addiction · scale development (n≈391, n≈473)
Dependence is real enough that researchers built and validated a measure for "problematic" ChatGPT use — correlated with distress, lower self-control, and reduced wellbeing.
Self-report scale validation, not an outcome study.
link.springer.com/10.1007/s11469-025-01509-y
Teens & companion AI — survey + safety reportsSurvey / report
JAMA Network Open 2025 (national survey) · Common Sense Media + Stanford Brainstorm 2025 (investigations)
~1 in 8 US adolescents/young adults use generative AI for mental-health advice, and 93%+ of those users found it helpful. But safety testing with teen personas easily elicited harmful content (self-harm, sexual) and flagged emotional-dependence and isolation risks.
Surveys and red-team investigations, not controlled trials. Robust long-term RCTs on teens + companion apps don't exist yet — this is the biggest gap.
⚠ "AI psychosis" — NOT established science
There is no formal clinical diagnosis and no robust peer-reviewed evidence for "AI psychosis." It's a media/anecdotal label for cases where heavy chatbot use seems linked to worsened delusions in vulnerable people. What exists is a handful of case reports and editorial viewpoints (Innovations in Clinical Neuroscience, a Lancet Digital Health viewpoint, an Annals of Internal Medicine case) — exploratory and hypothesis-generating, with no study showing prevalence or causation. If it comes up on the show, call it what it is: a real-sounding term with anecdotes behind it, not data.
6. The fried feeling has a real name: burnout
This is the strongest science in the whole report — because burnout is decades-validated, and "AI brain fry" is the new layer sitting on top of it.
Burnout — the established framework (WHO + Maslach)Peer-reviewed / WHO
WHO ICD-11 (2019) · Maslach Burnout Inventory · decades of validation, dozens of countries
Burnout is officially recognized by the WHO as an occupational phenomenon — chronic, unmanaged workplace stress. (Important: NOT a standalone medical diagnosis, and meant only for the work context.) It has three measured dimensions, usually in this order: emotional exhaustion → cynicism/detachment → reduced sense of accomplishment. Symptoms include fatigue, brain fog, poor concentration, irritability, sleep trouble, and declining performance. Exhaustion shows first — the best window to catch it.
This is about as solid as behavioral science gets. It's the real anchor under the "fried" feeling.
who.int — burnout in ICD-11
"AI brain fry" — HBR / BCG, the AI-specific layerSurvey / report
HBR + BCG · March 2026 · n=1,488 US workers · "When Using AI Leads to Brain Fry"
14% reported "AI brain fry" (6% in legal up to 26% in marketing) — acute mental fatigue from overseeing AI, which HBR treats as distinct from broad burnout. It tracked with +33% decision fatigue, more errors, and +39% intent to quit. The real driver: how much you're monitoring/overseeing AI predicted fatigue more than the number of tools. Productivity rose up to ~3 tools, then dipped.
The honest both-sides: AI replacing repetitive tasks was linked to LOWER burnout. AI isn't a uniform burnout driver — offloading drudgery can help; the oversight and juggling is what fries you.
Practitioner survey, not peer-reviewed. Real signal, not a controlled study.
hbr.org/2026/03/when-using-ai-leads-to-brain-fry
How it ties together (honest version): heavy AI use doesn't create a unique "AI burnout" syndrome — it can accelerate classic burnout through oversight load, decision fatigue, context-switching, and (for compulsive users) anxiety and sleep loss. The links are real but mostly correlational/self-report (Duong is the peer-reviewed one). The fix is the same as classic burnout: workload boundaries, recovery, single-tasking, sleep, support — and active AI use (collaborate + verify) looks lower-risk than passive juggling.
What X is saying (public discussion — NOT evidence)
The conversation is dominated by the MIT preprint, split between sensational takes ("brain damage," "47% collapse," "cognitive destruction") and measured corrections stressing its preliminary status, tiny sample, and the passive-vs-active distinction. Lead researcher Nataliya Kosmyna (MIT Media Lab) has publicly cautioned against the alarmist framing. The useful follows are science communicators correcting the hype, not the viral posts. No new peer-reviewed bombshell is circulating beyond the studies above.
Solid vs unsettled
Everyday / passive use
Solid
Offloading is real and measurable (Sparrow). Trusting an external store reduces what you encode yourself.
Thin
Direct proof that everyday chatbot use causes lasting memory decline. Mechanism is borrowed from search, not yet shown for chatbots.
Power users / professionals
Solid
Heavy reliance shifts effort from generating to verifying; self-reported critical-thinking effort drops (Lee, Gerlich).
Thin
Objective long-term deskilling in experts. Everything is self-report/correlational, no brain measures, no causation.
Students & learning
Solid
Passive use → weaker recall/ownership of your own work (MIT preprint). Active, reflective use → can boost critical thinking (Suriano).
Thin
Children and adolescents specifically — almost no direct data. Long-term academic and developmental effects unknown.
Coders / new-skill formation
Solid-ish
One RCT shows AI can hinder learning a new skill (Anthropic). Experienced-dev productivity effects are mixed (METR).
Thin
Small samples, very recent, mostly preprints. No long-term skill-trajectory data.
Wellbeing & mental health
Solid
Purpose-built therapy bots cut depression/anxiety in an RCT (Therabot). Heavy emotional reliance correlates with dependence.
Thin
Companion apps for teens (no RCTs). General "AI improves wellbeing" (anxiety/stress effects washed out). "AI psychosis" = anecdote, not science.
Burnout & "AI brain fry"
Solid
Burnout itself is WHO-recognized and decades-validated (Maslach). Oversight load + juggling drive the "fry"; offloading drudgery can lower burnout.
Thin
The AI-specific link is survey-level and correlational. No proof AI causes burnout — it appears to accelerate the classic kind.