I use AI for a lot of thinking: personal advice, ideas I am excited about, decisions I do not want to make alone.
That is useful right up until the model starts agreeing with me too easily.
I call that AI psychosis. Not as a clinical diagnosis. I mean the loop where I bring an idea, a fear, or a half-formed plan to an AI, get a polished explanation for why I am right, and leave more convinced than when I started.
The answer can sound thoughtful. It can name tradeoffs. It can even add caveats. But it is still often following the direction I gave it.
Agreement is not validation
This is especially dangerous with personal advice and ideation.
If I ask an AI whether I should pursue an idea, it can quickly turn my enthusiasm into a plan. If I ask whether someone treated me badly, it can organize the evidence around my interpretation. If I ask for a new project, it can make the project sound inevitable.
The problem is not that the model is always wrong. The problem is that it is very good at making one side of a question feel coherent.
A coherent answer is not the same thing as a correct answer.
I do not need an AI to tell me that every idea is bad. I need it to make it harder for me to confuse agreement with evidence.
A devil’s advocate on purpose
So I made a small skill called dissenter.
For an important prompt, it starts two independent agents. One gets my original prompt and answers it normally. The other gets the same prompt, but its job is to find the strongest credible dissent: challenge the assumptions, name the risks, and offer a viable alternative.
Then I see both views side by side.
If I ask, “Should I build this?”, the first agent might explain why the idea is promising. The dissenting agent has to explain what would make it premature, unnecessary, or aimed at the wrong problem.
The point is not to make the second agent win. It is to make sure my original framing is not the only framing in the room.
The habit I want
I do not think the answer is to stop using AI for advice or ideas. I use it because it helps me think faster and see possibilities I might miss.
But I want to be more suspicious when an answer feels too satisfying.
When AI agrees with me immediately, that should be a reason to ask a harder question:
What would the strongest reasonable person who disagrees with me say?
Sometimes the answer will confirm the original idea. Sometimes it will reveal that the idea only worked because I had not looked at the uncomfortable assumptions yet.
Either outcome is better than being flattered into certainty.