AI tends to tell you almost any decent idea is good. Ask what it thinks of your plan and it finds the strengths, smooths the doubts, and hands you back a slightly more confident version of what you already believed. That’s comforting. It’s also useless for judgment, and a little dangerous, because the more it agrees, the more sure you get.
So I stopped asking whether my thinking was right. I give it a different job now. Disagree with me, on purpose. It usually won’t on its own. Left to its defaults it leans toward telling me what I want to hear. Told to be hard on the weak spots, it starts to earn its keep.
A while back I built a small framework for mapping the people a decision touches. Who’s affected, who needs to be brought in early, who carries the risk if it goes sideways. I was proud of it. Before I used it for anything real, I ran it past AI and asked it to poke holes. It caught something I hadn’t wanted to see. My map was full of the agreeable people. The ones most likely to be hurt by the change, the ones who would resist for good reasons, were the ones I’d quietly left off. My own tool had a blind spot shaped exactly like the people I find hardest to face.
That’s the pattern I keep hitting, and it’s why I bother. What comes back is rarely a smarter analysis or a cleverer move. It’s a person I’d stopped seeing. I assumed context nobody else had. I filed someone under “on board” who wasn’t. I optimized for my own clarity and missed whose week it complicated.
There’s one discipline that makes this work. I write down my own read before I ask, so its answer sharpens my thinking instead of quietly becoming it.
The useful answers are almost never technical. They’re human. That’s how I know I asked the right way. When AI hands me a person instead of an answer, it’s doing the one thing I actually need it for.
If you want to try it, the whole thing takes about five minutes. Before you ask anything, write down your own read of the decision in a line or two, so you have your own thinking to protect. Then hand your plan to whatever AI you use and ask it to play the skeptic.
“Be hard on this, not kind. Assume I’m too close to it. Who does this hurt, who would resist for good reasons, and who carries the cost if I’m wrong? If you’re being agreeable, you’re not helping me.”
It won’t work every time. A lot of what comes back is generic, or invented, the same flattery just wearing a critical face, so you’re reading for the one answer that lands on someone real. Don’t ask whether the idea is good. Ask who’s missing.
When it hands you a name, go find that person before you decide, and ask what you’re not seeing. Sometimes you can’t, the call is a layoff, or something told to you in confidence, or the timing won’t wait. When that’s true, at least carry them as a real person with a real stake, not a name you optimized around. Either way, the AI only pointed. The rest is human.
If AI only ever makes me feel smart, I’ve used it wrong.