A word having a moment
In late 2024, Oxford University Press named "brain rot" its Word of the Year. More than 37,000 people voted. Usage of the term had risen 230% in twelve months. The definition Oxford gave is careful in a way the internet mostly ignored: the "supposed deterioration of a person's mental or intellectual state" from overconsuming trivial online content. That word "supposed" is doing a lot of work.
Two details from the announcement stuck with me. First, the term was coined for doomscrolling - low-quality content, endless feeds - and it was adopted most enthusiastically by the generations doing the scrolling. Second, sitting beside it on the same shortlist was another word: slop.
Keep those two together. They'll meet again in a minute.
Since then, brain rot has drifted onto AI. The claim, roughly: use AI heavily and your thinking decays. I get asked about it in training rooms fairly regularly now. And I used the term myself - until this week there was a card on my frameworks page called Brain Rot. I've renamed it, and this post is the reason why.
What the research actually shows
The short answer: something real, and less than the name suggests.
The strongest single finding I've seen comes from a large study of school students published in PNAS. Students given unrestricted GPT-4 access did better on practice problems while they had it. When it was taken away, they scored 17% lower on unaided maths than students who had never had access at all. A version of the tutor built with learning safeguards mostly removed the effect. Better with, worse without, design matters - that's the pattern in one sentence.
A Microsoft survey of 319 knowledge workers found that the more confidence people placed in AI, the less critical thinking they reported doing themselves - and that their thinking shifted towards checking and overseeing AI output rather than generating ideas. That's self-reported and correlational, so it can't tell you what causes what. A widely reported MIT experiment measured brain activity while people wrote essays with and without ChatGPT, and found lower engagement and weaker recall in the AI group - suggestive, but it's a small preprint that hasn't been through peer review yet.
And on the other side of the ledger: a systematic review of 69 experimental studies found ChatGPT generally improved academic performance and motivation. The picture is mixed, the effects are task-specific, and nobody has long-term data yet - these tools are too new.
So there is decent evidence that if you stop practising a skill because AI does it for you, your unaided performance on that skill drifts down. There is no evidence that AI use causes broad cognitive decline over years. One of those claims is "use it or lose it". The other is "rot". They are not the same claim.
Why the name bothers me
Rot is something that happens to you. Wood rots in the rain whether it likes it or not. But nothing in those studies describes damage being done to anyone. What they describe is practice that stopped happening. Mental arithmetic faded after calculators. Map-reading faded after GPS. Nobody's brain rotted; they stopped doing the thing, and the skill did what unpractised skills have always done.
The distinction matters because the two framings point to different responses. If AI rots your brain, the answer is abstinence and fear. If AI removes practice, the answer is deciding which practice you keep. One of those is useful.
The lazier truth
Here's my actual view, and it's less dramatic than either camp would like.
What I see in real businesses isn't decay. It's temptation. The model hands you a plausible answer in eight seconds, and the question is whether you do the work of judging it - or just take it. Take it often enough and two things happen at once. The output gets worse, because unverified AI output is how slop gets made (told you those two words would meet again). And your own judgement gets less exercise, because the judging was the practice.
I've written before about the Verification Tax - the checking time you have to spend before AI output is safe to use, and the true cost that decides whether an AI workflow is worth it. This whole conversation adds a second reason to pay it. Verification isn't just quality control on the machine. It's the remaining place in the workflow where your expertise still gets a workout. Skip the tax and you pay twice: slop now, softer judgement later.
Not rot. Laziness - or, said more kindly, a perfectly human preference for the easy answer, arriving at the exact moment the easy answer got very easy indeed.
What I'd actually do
The fix is not giving up AI, and anyone selling you that has a different problem with their reasoning. The fix is deliberate practice.
Decide which skills you will not let fade - the ones your clients actually pay you for - and keep doing some of that work unassisted, on purpose. Use AI as a sparring partner rather than a ghostwriter: have it attack your reasoning, find the holes in your plan, argue the other side. And on anything that matters, pay the verification tax in full, because that's where your judgement stays sharp.
The research term for all of this is cognitive offloading, and that's what my framework card is now called. It's a less catchy name than brain rot. It's also one I can defend to a sceptical solicitor without crossing my fingers, and in this line of work that's the better trade.
The word of the year tells you what people are afraid of. It doesn't tell you what's true. Those are different jobs, and only one of them is yours.
-- Alastair
P.S. The renamed card is here: Cognitive Offloading - one page, with the Pipeline Problem as its organisational twin.