gpt4-ece-rises-0007-to-0074-post-rlhf
IN premise — summaries/2026/08/24/kalai-2023-hallucination-inevitable-s2-related-work.md
Created 2026-08-24T17:10:59+00:00
GPT-4's Expected Calibration Error rises from 0.007 (pre-trained) to 0.074 (post-RL/PPO) on multiple-choice tasks, indicating that RLHF alignment reduces hallucination at the cost of increased calibration error.
Summary
The alignment training that makes GPT-4 less likely to confidently fabricate answers actually makes its self-reported confidence levels less reliable — if the model says it is 90% sure of a multiple-choice answer, its actual accuracy drifts further away from that 90%. This matters for any downstream system that uses the model's confidence scores to judge trustworthiness, because after alignment a "sure" no longer maps as cleanly to "right," so thresholds and risk decisions built on those scores need recalibrating.