How Expertise Gets Performed Instead of Proven
“It ain’t what you don’t know that gets you into trouble. It’s what you know for sure that just ain’t true.”
Have you ever heard that quote? It is attributed to Mark Twain. Funny thing. Mark Twain never said it.
Scholars at the Center for Mark Twain have found zero evidence he ever wrote or spoke those words. The closest source is probably Josh Billings, circa 1874. But folks have been confidently attributing it to Twain for decades, because it *sounds* like something Twain would say.
That’s right. A quote about the danger of false certainty, misattributed with total certainty.
Which brings me to something I read this week.
Researchers at the University of Southern California just published a study showing that when you start an AI prompt with “You’re an expert in….” — yep, one of the most common prompting tips on the internet — the AI actually gets *dumber*.
Not dumber-sounding. Dumber in fact. When tested on a standard knowledge benchmark, the “expert” prompt dropped accuracy from 71.6% to 68.0% across every subject category. The AI wasn’t thinking harder. It was performing confidence.
The researchers’ explanation: the expert persona activates the model’s instruction-following mode, which hijacks cognitive resources that would otherwise be used for factual recall. The AI spends so much effort sounding like an expert that it gets worse at being one.
But here’s where it gets interesting for anyone designing learning systems. The expert persona *did* improve something: alignment. The outputs sounded better. They were more polished, better structured, more “professional.” A human judge rating the responses would score them higher.
This is the difference between performance and competence. A learner who sounds fluent isn’t necessarily one who can transfer knowledge to a new situation. A presentation that looks polished isn’t necessarily one built on solid reasoning.
And the same thing is true for AI.
When the lead researcher was asked what this means for how people should prompt, his answer was interesting. When you care more about alignment, safety, rules, structure-following, be specific about your requirement. If you care more about accuracy and facts, do not add anything. Just send the query.
Just send the query. No flattery. No role-play.
The 85% of AI work, the part that’s already commoditized, is people copying prompting tips from the internet and hoping for magic. The 15% that actually matters? Understanding the cognitive architecture well enough to know when structure helps and when it hurts?
That’s not a prompting trick. That’s a design decision.