How do you know when you’re collaborating with AI, or outsourcing your judgment to it?

PART 2 — Metacognition is the skill. Here’s what the research says.

Last week I told you about the moment I almost lost my TEDx talk to an AI loop I couldn’t see. The question I left you with was the same one I’d been sitting with for months: how do you know when you’re collaborating with AI, or outsourcing your judgment to it?

And the capacity to know has a name: metacognition. And it’s also exactly what saved my TEDx.

You can’t know when you’re outsourcing your judgment unless you have judgment to begin with. And judgment is not a default. It’s a practice. Like every practice, it needs conditions.

Metacognition is not just ‘thinking about your thinking,’ that’s the shortcut definition, and it undersells it. Metacognition is a relationship with your own mind. And like any real relationship, it lives in everything at once: what you’ve lived, what you know, what you notice in the moment, how you interpret what’s happening, what you feel in your body before you’ve found words for it. Metacognition is what keeps you in contact with all of that while it’s happening. And judgment is what emerges from that contact. The stronger the relationship, the clearer the discernment. The clearer the discernment, the easier it is to notice when something outside you is quietly becoming your conclusion.

That’s where the research keeps landing. While better tools and AI literacy matter, they don’t remove the need for the human capacity to create the internal conditions under which good judgment can actually emerge. Metacognition is what creates those conditions.

Hear me out. You finished a task using AI, and the output looked good. You felt capable, maybe more capable than usual. What Fernandes and colleagues found is that this is exactly when the problem starts. People using AI performed better on reasoning tasks and, at the same time, became less accurate at knowing how well they’d performed. The more they used AI, the worse their self-assessment got because correct and incorrect increasingly feel the same from inside the experience. One way I make sense of that gap is this: AI makes the experience of arriving at an answer unusually fluent. And some of the friction we normally use to sense uncertainty disappears with it. The study doesn’t prove that mechanism.

This matters more than it might seem, because every interaction with AI contains delegation decisions. What do I hand over? What do I keep? When do I need help?
That’s the premise behind the Knowing Not to Know research. They found that how accurately people understood their own ability affected how well they made that call. When people were better calibrated about what they could and couldn’t do, they delegated to AI more effectively. When that self-reading was weaker, so was the delegation.
And delegation is only the first decision. Once AI gives you something back, another one begins: what do you do with it?

That’s where Soleimanof and Neufeld pick up the story. They found that metacognitive sensitivity shaped how consistently people responded to AI advice. People who could better distinguish when their own judgment was strong from when it was weak were better able to use that confidence as a signal for how much weight to give AI.

So metacognition matters on both sides of the interaction: when deciding what to hand over, and when deciding what to do with what comes back.

The instinct at this point is usually: okay, so let’s train people to use AI better. More literacy. Better prompting. More understanding of how the tool works. But Sidra and Mason found that AI literacy and AI metacognition are not the same thing. Knowing more about the tool doesn’t make you better at reading yourself. In fact, the Fernandes data suggests that higher AI literacy was actually associated with worse self-assessment accuracy. You can be an excellent AI user and still have no idea when it’s steering your judgment.

The next instinct is: fine, I’ll just ask AI to challenge me. Play devil’s advocate. Push back on my thinking. That’s a reasonable move, and it’s not enough. It still leaves you with the harder problem: judging the challenge. AI can generate ten strong counterarguments on command. That doesn’t tell you whether those counterarguments matter, whether they should change your conclusion, or whether they’re simply plausible-sounding alternatives. Someone prone to self-doubt may treat every counterargument as proof their idea is weak. Someone very confident may dismiss them all. What you do with the challenge depends entirely on your relationship with your own mind in that moment. You can’t tool your way out of a human challenge and responsibility.

So what’s the actual variable?

Soleimanof and Neufeld found that the people who made better calls with AI weren’t smarter or more experienced. They had a more accurate and sensitive relationship with their own cognition. The ones with lower metacognitive sensitivity showed a very specific pattern: sometimes overriding correct judgment, sometimes sticking to weaker judgment. Not random noise. They couldn’t read themselves accurately enough to know which situation they were in. And AI’s mechanism is partly why.

AI tends toward legibility, human thought does not.
“That’s not X, that’s Y”. That’s the move AI does better than anything: take a messy, ambiguous thought and return it clean.
We know that rhythm by now. But human judgment usually lives before that clarity, in the messy, embodied, not yet resolved place where the real thinking happens. If you’re not in relationship with your own mind while AI participates in your thinking, you won’t notice when you’ve handed that place over.

Which brings us back to the original question. You are collaborating when AI contributes to the thinking and you remain able to locate your own role in the frame, interpretation and decision. You are outsourcing judgment when those decisions are being made somewhere in the process, but you can no longer tell which ones you actually made, or consciously chose to hand over.

What we do shapes the future, and what we do is shaped by how we think. If you’re not paying attention to your own mind, something else is doing it for you.

If you want to strengthen your team’s judgement, Own Your Thinking is a one-day intensive for teams who want to stay the author of their decisions as AI becomes part of the process.

Know more about Own Your Thinking solution.