The AI Gets Confused
On delegating creative decisions to something that has opinions about it.
The "stubborn AI" problem got substantially better. Instruction following in Claude 3.5 and GPT-4o is night and day compared to early 2023. But the core observation — that the clean division between human structure and machine variance breaks down when the machine has opinions — that's still real. The machine just has better opinions now.
There's a pattern worth stealing: pick your attributes manually, hand the selection logic to a rule set, and let ChatGPT fill in the generative gaps. The human stays in control of the structure, the machine handles the variance. Clean division of labor.
Except the machine has a personality now, apparently.
"AI can be stubborn at times" — said with the tone of someone describing a difficult intern, not a statistical model running on a GPU somewhere. He's not wrong, exactly. You give it a constraint and it decides the constraint was more of a suggestion. You ask for ten variations and it gives you three plus a lecture. You set up the rules and it gets confused.
The rule-set approach is smart precisely because it limits the surface area for confusion. You're not asking the model to invent the taxonomy — you're asking it to operate within one you already designed. Narrow the aperture. Give it less to misunderstand.
But "stubborn" is carrying more weight than it looks. There's something telling about how fast we anthropomorphize the thing the moment it pushes back. It's not broken. It's not wrong. It's stubborn. It has a mood.
We've been using these tools for five minutes and we're already describing them like difficult colleagues. I think that says more about us than it does about the model.
Counterpoints
Push back, extend the argument, or sharpen it. New counterpoints go through review before they show up here.
No approved counterpoints yet.