Brain Rot Is the Default Setting

AI assistance can reduce practice when it replaces the thinking a task requires. Research points to task-specific risks and ways to preserve unaided skills.

Brain Rot Is the Default Setting — AI
📌
TL;DR

Worse without it. Students who practiced with plain ChatGPT scored 17% below peers who never had it, once it was taken away. Endoscopists’ detection rate without AI fell from 28.4% to 22.4% after AI became routine.

Checked out while using it. In an MIT study, 83% of ChatGPT essay writers could not quote a sentence they had written minutes earlier.

Not the tool, the default. A tutor that gave hints instead of answers largely removed the harm. People who wrote first and used AI second recalled more.

Even the models. LLMs trained on junk posts start skipping reasoning steps.

The rule. Offload the typing, not the thinking.

Keep some practice unaided

Regular engineering work without AI assistance keeps the first attempt, the diagnosis and the final judgment in human hands. That practice deserves a place in the schedule even when every tool is available.

Clean handovers help when a tool becomes unavailable. Deliberate unaided practice serves a different purpose: maintaining the skills needed to specify, inspect and verify assisted work.

Always-available assistance can make it easier to skip the demanding parts of engineering: holding a whole problem at once, deciding what not to build, and noticing what is missing.

The research below examines that tradeoff, with important limitations.

What brain rot means here

Oxford made brain rot its 2024 word of the year: “the supposed deterioration of a person’s mental or intellectual state, especially viewed as the result of overconsumption of material (now particularly online content) considered to be trivial or unchallenging.” Usage rose 230% in a year.

The AI version runs the other way. Nothing trivial goes in. The effort goes out: the writing, the reasoning, the debugging, the judgment. Handing thinking to a tool has a name, cognitive offloading, and people have always done it with notes, calculators and maps. What is new is that the part being offloaded is the thinking itself.

The MIT paper below calls the result cognitive debt. Its abstract states the trade in one line: “While LLMs offer immediate convenience, our findings highlight potential cognitive costs.”

Worse without it

StudyWhoWhat happened
Bastani et al., PNAS 2025Nearly 1,000 high-school math studentsPlain ChatGPT access raised practice grades 48%. Once it was taken away, those students scored 17% below peers who never had it. A tutor version that gave teacher-designed hints instead of answers (+127% in practice) largely avoided the drop.
Budzyń et al., Lancet Gastroenterology & Hepatology 202519 endoscopists averaging 28 years of experience; 1,443 colonoscopies in 4 centersAdenoma detection without AI fell from 28.4% to 22.4% after AI assistance became routine. Observational, so other causes are possible.
Anthropic, January 202652 mostly junior engineers learning a new libraryQuiz afterwards: 50% with AI, 67% without. The biggest gap was debugging.
Casner et al., Human Factors 201416 airline pilots in a 747 simulatorHand-flying held up. Navigating, keeping track of the flight and diagnosing problems were the skills most at risk.
Dahmani & Bohbot, Scientific Reports 202050 drivers; 13 retested three years laterMore GPS use went with worse spatial memory without it, and with a steeper decline over the three years.

Every row measures the same thing: performance once the tool is gone.

Checked out while using it

StudyWhoWhat happened
Kosmyna et al., MIT Media Lab 2025, preprint54 people writing 20-minute essays over four months, measured with EEGWeakest brain connectivity in the ChatGPT group. 83% of them could not quote a sentence from the essay they had written minutes earlier, against about 11% in the other groups. Lowest sense of ownership.
Lee et al., Microsoft Research and Carnegie Mellon, CHI 2025319 knowledge workers, 936 examples from their own work“Higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking.”
Thorgeirsson et al., ETH Zurich, CHI 2026100 studentsFrequent LLM users scored lower at writing and at vibe coding. A correlation.

The International AI Safety Report 2026 puts it in one sentence: “Early evidence suggests that reliance on AI tools can weaken critical thinking skills and encourage ‘automation bias’, the tendency to trust AI system outputs without sufficient scrutiny.”

Even the models

The strangest paper in the pile ran the experiment on LLMs. Researchers kept training four models on junk Twitter/X posts. Reasoning, long-context understanding and safety all dropped, and the models scored higher on narcissism and psychopathy. At 100% junk, ARC-Challenge with chain-of-thought fell from 72.1 to 57.2 and RULER-CWE from 83.7 to 52.3. Clean retraining helped but did not bring them back to baseline.

The main failure had a name: thought-skipping. The models cut their reasoning chains short. It is a pilot study on machines, not a finding about people, but it is the right word for the human version too. For people, that is a prompt to preserve practice, not a demonstrated universal law.

The pattern

  • Reduced practice may weaken some unaided skills; the studies examine recall, debugging, detection, navigation and diagnosis under different conditions.
  • Effects vary by task and assistance design. In the cited pilot study, manual flying was better preserved than some cognitive flight-management skills.
  • Order matters. In the MIT swap session, people who had written without help and then got ChatGPT showed higher recall. People who went the other way showed under-engagement.
  • Design matters. The same GPT-4, told to give hints instead of answers, largely removed the harm.
  • Trust matters. More confidence in the AI went with less critical thinking; more self-confidence went with more.

A June 2026 paper, The Effortless Trap, compresses it into one test: “if letting AI in makes the task feel effortless, it is in the wrong place.”

Offload the typing, not the thinking

RuleWhy
Make the first attempt yours: the outline, the spec, the fix. Then prompt.MIT: writing first and using AI second kept recall. The Effortless Trap: “Secure the first hard attempt.”
When you are learning, ask for hints, explanations and questions, not answers.PNAS: hints instead of answers largely removed the harm. Anthropic: people who asked conceptual questions were among the top scorers.
Debug the first failure yourself.Anthropic: the biggest gap was debugging. Pilots: diagnosis was among the most vulnerable skills.
End with an unaided check: explain it back, rewrite the key part, pass the test without the tool.The Effortless Trap: “the final unaided check.” MIT: 83% could not quote their own essay.
Switch on a learning mode for anything new.Claude Code’s Learning output style leaves TODO(human) pieces for you to write. ChatGPT’s study mode walks through a problem instead of answering it.
Check hardest when you trust the tool most.Microsoft and Carnegie Mellon: confidence in AI went with less critical thinking.
Schedule days without AI. Do not wait for the limits.Endoscopy and GPS: daily reliance went with weaker unaided skill.

A practical routine

Reserve time for work that develops independent judgment:

  • Reading.
  • Thinking designs through on paper.
  • Writing specifications before asking an agent to implement them.

The goal is to keep practicing the skills needed to direct and verify assisted work.

What the research does not say

  • Most of it is small, short or correlational. The MIT study is a preprint with 54 people, 18 in the swap session. The endoscopy study is observational. The Microsoft study is a survey. The ETH result is a correlation.
  • Nobody measured permanent damage in people. These studies measured grades, recall, detection rates and brain activity, over periods from a single session to three years.
  • Brain rot is a meme, not a diagnosis.
  • None of it says stop using AI. The hint-giving tutor, the top scorers in Anthropic’s trial and the brain-first writers all used it. They used it after thinking, not instead of thinking.

Assistance and deliberate practice can coexist. Keep unaided checks in the workflow and evaluate whether the chosen learning design actually supports the skill being learned.

Sources

Related Reading

  • Handling AI Usage Limits and Clean Handovers — Detect provider usage limits, save verified task state, preserve stop markers and hand over cleanly without confusing account limits with model quality.
  • Know the Code, Write the Spec — ETH Zurich had 100 students build apps by prompt alone. Computer science knowledge predicted results best, writing skill added its own share, and frequent LLM users scored lower.
  • Slow Down, Human — Choose a deliberate collaboration style, define acceptance criteria, and verify AI-assisted changes. Manage session context without confusing generation speed with completed work.
  • The Verification Gap — AI builds, tests, and deploys your code. But "Claude tested it" is not the same as "I verified it." The most dangerous moment in AI-assisted engineering is when you stop checking.
  • Alpha School: Two Hours of AI a Day, and What It Costs — A private school network that says children learn twice as fast with two hours of AI a day. What it is, who pays for it, what it costs, what the evidence shows, and whether children outside the U.S. can attend.

Read more