kalai-two-hallucination-regimes
IN premise — summaries/2026/08/24/kalai-2023-hallucination-inevitable-s0-abstract.md
Created 2026-08-24T17:10:58+00:00
Singleton/arbitrary facts have an unavoidable hallucination lower bound for calibrated models, while repeated/systematic facts (e.g., publication references, arithmetic rules) have no statistical lower bound forcing hallucination
Summary
Even a perfectly well-calibrated model will occasionally invent details about rare, one-off facts — that error floor is mathematically unavoidable. But for structured, recurring patterns like citation formats or arithmetic rules, there is no such floor, meaning a good model can get arbitrarily close to zero confabulation on those. The practical upshot is that you should build different trust and verification strategies depending on whether a given fact type is idiosyncratic or systematic.