24 October 2026AI

I haven’t had a free thought in two years

A sentence that didn’t exist

I overheard two students in the university library last week. Sharp ones. They were having the kind of mature, considered conversation you’d happily be part of. Then one of them said, half laughing, I haven’t had a free thought in two years.

He was joking. He has obviously had thoughts. But he recognised something in himself, said it out loud, and nobody at the table found it strange.

Five years ago that sentence could not have been said. Not because nobody felt distracted, but because there was nothing to hand your thoughts to. These students arrived at university in the same term as ChatGPT. For every essay they’ve ever been set, something else could have written it for them. That has never been true of any student before.

I use these tools constantly, on my own time, for all sorts of small things: a YouTube channel I produce with no video editor, little personal helper tools like a watch app for catching ideas and a dashboard for organising my own workflows, my book Turn Down the Tap, and this piece, which started as me dictating into my phone on a walk. So this isn’t a piece about AI being bad. It’s a piece about where the thinking goes.

The best document you’ve ever seen

Here’s the same thing from the other end of a working life, and this time it’s the research talking, not me.

Researchers at BetterUp Labs and Stanford have a name for what they found.

In their survey of 1,150 US employees, four in ten had received some in the past month. Each piece cost the person on the receiving end nearly two hours to sort out, and about half of them thought less of the colleague who sent it afterwards.

The pattern they describe goes like this. Someone asks the AI for a plan. What comes back is the most impressive document they’ve ever had on their screen. Structured, confident, complete. So they read it once, feel the glow, and forward it on.

Nobody checks it. The sender doesn’t, because it looks finished. The people receiving it don’t either, because it looks finished to them too. And then it generates meetings.

The detail I find most telling is their description of what qualifies: content that’s unhelpful, incomplete, or missing context about the project. Not wrong, exactly. Just not for anyone in particular.

What a thorough machine does with a thin prompt

The machine isn’t failing here. It’s doing exactly what it was asked.

Ask it for everything you might need for a job and it will give you everything. It’s thorough; that’s its nature. If you didn’t tell it who the audience is, what the budget is, what’s already in place and what’s already been ruled out, it will hand back a list that asks the world of whoever has to act on it. A perfectly reasonable document for a company that doesn’t exist.

The gap between a useful document and workslop is almost never the model. It’s what went in before the prompt and what happened after the answer.

Pilots and passengers

The same researchers split people into pilots and passengers. Pilots use AI to extend their own thinking. Passengers use it to avoid thinking.

I’d draw the line somewhere slightly different: whether you had to think before the tool existed.

Most adults have twenty or thirty years of figuring things out without it. We have judgement, built the slow way, by getting things wrong and having to live with it. That judgement is what makes the tool useful. You can tell when the answer is too generic, when it’s missed the obvious, when it’s being agreeable rather than right.

That’s the pilot’s advantage, and it wasn’t earned on the tool. It was earned before it.

How I actually use it

For what it’s worth, here’s what my own use looks like, on my own stuff: the channel, the little helper tools, the book, this site.

I start by telling it what I know. Here’s my understanding of the problem, here’s what I’ve already tried, here’s where I think the risk is. Then I send it off: go and look at what’s out there, see if there’s a cheaper or more secure way of hosting this little personal tool, tell me what I’ve missed. It’s a research assistant with a very long reading list.

When it comes back, I argue with it. It still doesn’t have the full picture, so I give it more and send it round again. I ask it to make the case against the thing I’ve just decided. I ask for three alternatives before I pick one.

The thing to watch is that it wants to please you. It will agree with you if you let it, and you have to be deliberately awkward to stop it: don’t tell it what you hope the answer is, ask it where you’re wrong, and treat a quick yes with suspicion.

Used like that it isn’t doing my work. It’s the advisor, or the room full of advisors, that most people never had. No author ever wrote a book alone; they had editors and pubs and people to argue with over dinner. If you were unlucky enough not to have that, you were stuck. Now you’re not. That’s a whole piece on its own, so I’ll leave it there for now.

Back to the library

Now the other end: people who got the tool before they got the judgement.

The evidence here is early, and I’d be a hypocrite to overstate it in a piece about checking what you’re handed. But so far it points one way.

An MIT Media Lab study put 54 students on EEG while they wrote essays, some with ChatGPT, some with a search engine, some with nothing but their own heads. The ChatGPT group showed the weakest brain connectivity and could barely quote from essays they’d written minutes earlier. When they later wrote without it, the weaker pattern was still there. The researchers called it cognitive debt. It’s a small study, essay-only, not yet peer reviewed, and only 18 people did the final session. A flag, not a verdict.

A larger 2025 survey by Michael Gerlich found a strong negative link between heavy AI use and critical thinking scores, with 17 to 25 year olds the most dependent on the tool and the lowest scoring.

Same tool. The difference was whether they treated it as a conversation or a vending machine.

The more interesting finding is the one that splits the group. Students who used AI as something to question and argue with held on to their critical thinking. Students who accepted what it said showed little or none. Same tool. The difference was whether they treated it as a conversation or a vending machine.

So the student in the library isn’t doomed. He’s just missing the twenty years of friction the rest of us got for free.

Fig. 1 · Two questions The pilothas the judgement,and checksThe arguingstudentno judgement yet,checks anywayWorkslophas the judgement,doesn't checkNo free thoughtin two yearsneitherCHECKS WHAT COMES BACKyesnoyesnoBUILT JUDGEMENT BEFORE THE TOOL
The document was never the work.

The document was never the work

Here’s the thread that joins the two.

The document was never the work. The thinking before the prompt and the arguing after the answer, that’s the work. The document is the residue.

Workslop is what happens when someone who has judgement skips that step because the output looked finished. The student in the library is what happens when someone never built the judgement in the first place. Same failure, both ends of a career.

The tools will get better at giving you less. They’ll learn to ask what you meant before they answer. But right now, in this window, they’re generous to a fault, and the only thing stopping one from generating work for everyone around you is you.

Tell it what you know. Argue with what it gives you. And every so often, have a free thought, just to check you still can.

Sources

All musings