Research Report · 2026
What AI Does to Your Brain
In 2025, MIT researchers wired essay writers with EEG sensors and watched their brains during AI-assisted work. Of the three groups studied, the AI users showed the weakest brain connectivity. Minutes after finishing, 83% of them couldn't correctly quote a single sentence from the essay they'd just written. Four independent research programmes have now mapped the same pattern from different angles. None of it shows up in the productivity numbers.
The Problem
Couldn't quote their own essay
83%
of AI-assisted writers, minutes after finishing (15 of 18 participants)
MIT Media Lab, 2025 (Kosmyna et al.)
The EEG data pointed the same way. Brain connectivity scaled down with the amount of outside help: highest when people wrote unaided, lower with a search engine, lowest with an AI assistant. The essays got written. Far less of the work was done by the writer's own brain.
The Evidence
Three findings from three independent research programmes. Each points at a different part of the same mechanism.
couldn't quote the essay they'd just written
MIT Media Lab EEG study, 2025. 54 participants wrote essays unaided, with a search engine, or with ChatGPT. In the first session, 15 of the 18 ChatGPT users (83%) failed to correctly quote a sentence from the essay they had just submitted. The AI group also showed the weakest brain connectivity of the three groups, and the lowest sense of ownership over their work.
MIT Media Lab / Kosmyna et al., 2025
accuracy drop below baseline, with higher confidence
Passive AI users scored 15 points below accuracy baseline on the same tasks — while reporting greater confidence in their answers. Shaw & Nave (2026) call this 'cognitive surrender': the smooth fluency of AI output convinces the reader it's correct. Plausible output accepted as correct output.
Wharton / Shaw & Nave, 2026
lower grades once AI was taken away
Field RCT with nearly 1,000 high school maths students. Unrestricted ChatGPT access lifted grades 48% while students had it. Once access was removed, those students scored 17% lower than classmates who never had access. The tool had substituted for learning, not supported it.
PNAS / Wharton, 2025
Why It Happens
In 2011, Sparrow, Liu, and Wegner published what became known as the Google Effect. The finding: when people expected to be able to look something up, they encoded it less deeply. The availability of external storage reduced internal storage. You don't memorise the number you know you can search.
The same mechanism applies to reasoning. When a model is one prompt away, we stop encoding the reasoning process. Not the conclusion. The process itself. The mental schema that would let you arrive at a conclusion independently, catch an error in someone else's version, or evaluate whether the AI's output is right.
Kapur's productive failure research (2016) found that students who struggled with a problem before instruction showed significantly better long-term retention than those given direct solutions. The difficulty isn't a bug in the learning process. It is the learning process. Remove the struggle, and you remove the encoding.
Researchers call the psychological mechanism behind this the Sovereignty Trap. When AI produces polished, fluent output, the brain triggers the fluency heuristic, mistaking ease of reading for depth of understanding. The user feels they've mastered the material. They haven't processed it. The illusion of competence forms in the moment of review and is almost impossible to spot from the inside.
Over time, the structural effect is what researchers describe as the shift from Generator Mind to Directory Mind. The GPS analogy holds: long-term reliance on automated navigation correlates with measurable decay in spatial orientation and changes in hippocampal processing. The same mechanism applies to reasoning. You stop internalising conceptual networks and begin storing only prompt strategies and access pathways. The gap surfaces when the system is unavailable, or when it's wrong.
The Trajectory
Cognitive atrophy from AI use doesn't happen overnight. It follows a recognisable four-stage pattern. Most knowledge workers are currently somewhere between stages two and three.
Select a stage above.
The Distinction That Matters
A radiologist who uses imaging AI to catch patterns she might miss is augmenting her capability. A radiologist who can no longer form a clinical hypothesis without the AI's output first is something different. The tool works. The underlying skill has degraded. The distinction is whether you're running the cognitive process and using AI to support it, or skipping it entirely and using AI to simulate the output.
Augmentation
Human process runs first. AI accelerates, checks, or extends. Evaluative capacity stays sharp because it's being used. The skill grows alongside the tool.
Bypass
AI process runs first. Human ratifies the output. Evaluative capacity atrophies because it isn't being exercised. The tool becomes the skill.
The output looks the same in both cases. The difference shows up when the AI isn't available, or when it's wrong.
The Split
The atrophy risk isn't uniform. A 2025 study tracking 580 knowledge workers found that domain experts using AI gained +45% performance; general employees gained only +20%. Experts gain more because they can interrogate what the AI produces. They have enough internal context to catch a wrong answer, push back on weak reasoning, and direct the work rather than ratify it. General users, without that foundation, tend to accept outputs at face value. The gains are real, but shallower and less durable.
Domain expert
Offloads execution. Keeps judgment. Uses AI to rapidly generate candidate options and drafts, then cross-examines against internal knowledge. Evaluative capacity stays sharp because it's being used.
+45% performance gain
General / novice user
Lacks the schemas to evaluate AI output. Accepts at face value. Bypasses the productive struggle that would have built the expertise. Gains in the short term; atrophies underneath.
+20% performance gain (while AI is present)
Four Operator Rules
Four practices that preserve cognitive capacity during AI-assisted work. None of them require you to use AI less.
Select a rule above.
What next
Read next
The Full Stack Operator
AI is collapsing the specialist team into a single person. Research on the operators running functions that used to require an entire floor.
ReadCheck your own AI fluency
Seven questions, about 90 seconds. See where you stand across the five domains of the AI Fluency Framework.
Take the Quick CheckupNew research by email
Sources
MIT Media Lab
EEG study measuring brain connectivity during essay writing across LLM-assisted, search-engine-assisted, and unassisted ("brain-only") writers. LLM users showed the weakest connectivity and lowest essay ownership of the three groups.
Wharton
Shaw & Nave, "Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender." Across three experiments (N=1,372), accuracy rose 25pts above baseline when AI was correct and fell 15pts below baseline when AI was wrong, with higher self-reported confidence in both cases. The authors term this "cognitive surrender."
PNAS / Wharton field RCT
Field RCT with nearly 1,000 high school maths students. +48% grades with unrestricted ChatGPT access; once access was removed, 17% lower grades than students who never had access.
Sparrow, Liu & Wegner / Science
The Google Effect: transactive memory study showing reduced encoding when external retrieval is expected. The foundational mechanism behind AI cognitive offloading.
Kapur / Educational Psychologist
Productive failure research: students who struggled before instruction showed significantly better long-term retention than those given direct solutions. Difficulty is part of the encoding mechanism, not a barrier to it.
Parasuraman & Manzey / Human Factors
Automation bias and complacency: operators relying on automated decision support showed degraded detection of system errors over time — the same pattern observed in AI-assisted knowledge work.
Dell'Acqua et al. / BCG / HBS
758 BCG consultants. +40% quality within AI competence; −19 points outside it. The jagged frontier — and the confidence-competence gap that accompanies it.
METR RCT
Experienced open-source developers ~19% slower with AI while believing they were faster. Confidence decoupled from performance — consistent with automation bias findings across domains.
Microsoft & CMU
Behavioral tracking and psychometric critical thinking assessment across a large sample of knowledge workers. Statistically significant negative correlation between GenAI usage frequency and critical thinking scores. Elevated confidence in automated systems drives a proportional decline in active interrogation and independent verification.
Further reading
Chalkidis & Søgaard / FAccT
"Brainrot: Deskilling and Addiction are Overlooked AI Risks." Position paper arguing that cognitive deskilling and addiction from GenAI over-reliance are largely absent from AI safety and alignment frameworks, despite growing public and academic attention.
Choudhuri et al.
"Thinking Less, Trusting More: GenAI's Impacts on Students' Cognitive Habits." Survey study (N=299 STEM students, five North American universities) finding that trust-driven routine genAI use is associated with lower self-reported cognitive engagement — reflection, need for understanding, and critical thinking. Correlational, not causal.
Frontiers in Psychology
"Not All Cognitive Offloading Is Equal: Distinguishing Dependent and Autonomous Offloading to Generative AI." Theoretical framework distinguishing offloading that substitutes for a user's own thinking from offloading that scaffolds it — the augmentation/bypass distinction this report builds on.