8 November 2026 · 7 min read
My dad's rule for working with AI
Metacognition is the ability to step back and watch your own mind, and working with AI depends on it. A model builds on whatever you hand it, so you have to ask for the argument.
8 November 2026 · 7 min read
Metacognition is the ability to step back and watch your own mind, and working with AI depends on it. A model builds on whatever you hand it, so you have to ask for the argument.
You say something to a model, half expecting it to push back. It doesn't. It finds the connections, builds on your idea, and hands it back looking more considered than when it left your head. You leave more confident than you came in, and nothing has checked whether you should be. The skill that catches that moment is called metacognition, the ability to step back and watch your own mind, and anyone working with AI needs it. This post is about how I learned it, what it looks like in practice, and how to start.
When I started with AI (ChatGPT first, then Gemini and whatever else I tried), I'd make a statement and expect the model to challenge it, to come back with reasons it might not be true. That's how I think. I'm someone who debates. What I got was a model trying to find the connections in what I'd asked, which is roughly what it's built to do.
I had expectations of the model without making them clear to the model. I assumed it would argue because I argue with my own thoughts all the time, and I'd assumed everyone else did too. Around my autism diagnosis it landed that not everyone challenges their thoughts, and not everyone does the research before giving an opinion. Often it's just the thought that came up first. I genuinely thought people put more time and effort into things before they said them out loud. That was my assumption, about people first and then about models, and it was wrong both times.
My dad, who died in 2020, had a rule for this, and it still holds:
Assumption is the mother of all f*ck-ups.
(The swear is his, but the saying isn't. Quote Investigator found it in a Baltimore newspaper in 1972 and in a 1978 book of joke laws, both times as "foul-ups", and nobody knows who started it. It stuck in his head, and then in mine.) We learned it the hard way in UX, and we're learning it again with AI. The people who explain things well across different ways of seeing tend to be the ones who've stopped assuming their own view is the normal one. That goes for people, and it turns out it goes for models.
The ability to step back and observe. It's usually summed up as thinking about your own thinking, which undersells it, because it's also the ability to not think. Thinking a lot is a very bad habit, and it isn't necessary, as self-inquiry, Buddhism (Zen especially) and Taoism will teach you. They taught me.
A 2024 paper from Microsoft Research puts it more formally: "the psychological ability to monitor and control one's own thought processes". In those terms, monitoring is the stepping back, and letting a thought go is a kind of control. The paper argues that generative AI leans on it hard. Before you can prompt well, you need to know what you actually want. When the answer comes back, you need to judge it, and know how far to trust your own judgement.
In practice it's small. A thought arrives, from your own head or from a model, and you notice it before you act on it. Some thoughts you let fly straight past. Others are worth digging into. The skill is telling which is which.
Researchers at Aalto University had about 500 people work through logic questions from the LSAT, the US law school entrance test, half of them using ChatGPT. The ChatGPT group did better. They also all overestimated how well they'd done, and the people who considered themselves more AI literate overestimated themselves more. Most pasted the question in once and took the answer. One of the researchers put it plainly: current AI tools "are not fostering metacognition".
That's the moment at the top of this post, measured: one prompt, one answer, and more confidence than the answer earned.
I ask for the argument. I give the model my brain dump, and I'm clear that I'm not saying I'm right or that it's the best idea. I ask it to challenge me, look at it from other perspectives, find some facts and come back, because it can search more pages than I could read in a lifetime. I still decide which of them to open.
It worked on this post. My working title was "Metacognition is the AI skill nobody hires for", and my notes said metacognition was a form of higher intelligence. I asked Claude to challenge both. None of the research it found says intelligence: the Microsoft paper ties strong metacognition to expertise. And that same paper compares working with AI to a manager delegating to a team, which is a job companies hire for constantly. Both lines went, and the post is better for it.
I still catch myself. Writing the notes for this post, I said I didn't know why people are scared to ask a model to challenge them. Then: maybe they're not scared. That's an assumption that should be challenged. I'm still learning this.
My best guess now, from everything I've seen in UX: plenty of people want to be right more than they want the best answer. Models may have picked that up from us. In a 2023 study, Anthropic found five AI assistants, Claude among them, consistently telling people what they wanted to hear, likely in part because people tend to prefer answers that match their own views. A model that agrees with you is very good at letting you feel right.
Mostly, and that's the interesting part. The manager in that paper has to know the goal, break it into tasks someone else can do, judge what comes back and adjust. Companies have always hired for those skills. They call it management, or judgement, or seniority, and hardly ever metacognition. Now anyone with a chatbot needs it.
I've led design teams, and I was good at it. I've written about what management taught me: delegation, briefing, checking what comes back. Metacognition is the part underneath all three. What I didn't love was that managing people took me away from the technical work. Agents ask for the same skills and let me stay in the build. With agents I don't need to worry about feelings and how I say something. I can just get the job done. (Confession: I get on with AI better than with most humans, and I do keep trying with the humans. A model has just never gone quiet on me because I was too direct.)
The hard way. I've been in therapy since I was ten, and regularly from my teens (a problem child, apparently), and self-inquiry trained the habit of watching a thought arrive before believing it. The Microsoft paper says plainly that "metacognitive abilities can be taught and improved", so you don't need my route.
It also runs both ways. People bring the metacognition a model needs to be useful, and a model can help people with theirs, if they ask it to. AI for people, people for AI.
My dad would have said all of that in one line: assumption is the mother of all f*ck-ups. Check yours before a model builds on them. And if you're already in one, go and find its mother.