The moment I catch myself getting defensive with an AI’s answer instead of checking it is the moment I have stopped thinking and started performing being right.
There is a specific feeling I have learned to treat as a warning light: the model gives me an answer I do not like, and instead of checking whether it is correct, I start composing a rebuttal. I am not evaluating evidence at that point. I am defending a position I had already taken before I asked the question, and the model’s answer is just the thing I am now arguing against. The tell is not that I disagree — disagreement is often correct — it is the order of operations: verdict first, evidence search second, when the honest order runs the other way.
It took me a long time to notice this pattern because it disguises itself as rigour. Pushing back on an AI’s output looks, from the outside, exactly like the healthy scepticism everyone tells you to have. The difference is entirely about direction. Scepticism starts from the question "is this true" and follows the evidence wherever it goes. What I am describing starts from "I need this to be wrong" and goes looking for the evidence to support that, which is a different operation wearing the same clothes.
The tell, once I started watching for it, is speed. A genuine check of a claim takes time — I have to go look at the thing the model is talking about, run the command, read the file, reproduce the behaviour. A defensive rebuttal is instant, because I am not gathering anything new, I am just retrieving reasons I already had lying around. If I notice myself typing a reply before I have opened anything, that is the signal, not the content of what I am about to type. This is uncomfortable to sit with in the moment, because the urge to respond immediately feels like engagement rather than avoidance, and slowing down enough to notice the gap between the two takes active effort every single time.
What actually regulates this for me is a rule rather than a feeling, because feelings are exactly the thing under question: I do not get to disagree with a tool’s output until I have produced one piece of evidence it did not have. Not a reason, not an intuition — a command run, a file read, a test executed. If I cannot produce that, the honest position is that I do not yet have grounds to disagree, whatever my instinct says.
This generalises past AI, obviously — it is the same discipline good engineers already had towards a colleague’s code review, or towards a second opinion from someone senior. What working with models constantly did was raise the frequency. I am now in this exact situation dozens of times a day rather than a few times a month, and something that happens dozens of times a day is worth having an actual rule for instead of relying on being in a good mood when it comes up. The volume changes the stakes too — a habit that costs me almost nothing once a week compounds into something worth actually fixing once it is happening before lunch most days.
The uncomfortable part is admitting how often the answer, once checked, was simply right, and my objection was never about the evidence at all. It was about not wanting to have missed something, or not wanting to redo work I had already convinced myself was finished. Neither of those is a reason a claim is false, and treating them as if they were is the exact substitution the rule exists to catch.
I would rather know that about myself in the moment than after I have shipped the wrong fix on the strength of an argument I never actually tested. The rule costs me almost nothing when the model is wrong, because then the evidence is easy to find. It only ever costs me the small discomfort of finding out I was the one who was not thinking straight, which is a cheap price for not finding it out later.