Story perspectives

OpenAI Aims to Curb AI Hallucinations with New Strategies

9/7/2025

44 10

Summary

Story summary

OpenAI's research shows language models hallucinate because training rewards guessing over acknowledging uncertainty.
The ChatGPT "Model Behavior" team has merged with the Post-Training team, led by Joanne Jang, to enhance human-AI interaction.
Current evaluation systems penalize uncertainty, prompting models to give confident yet incorrect responses.
Proposed changes include redesigning benchmarks to reward uncertainty and reduce guessing.
OpenAI seeks to improve AI reliability by adjusting evaluation incentives.

1 of 1

Story summary

OpenAI's research shows language models hallucinate because training rewards guessing over acknowledging uncertainty.
The ChatGPT "Model Behavior" team has merged with the Post-Training team, led by Joanne Jang, to enhance human-AI interaction.
Current evaluation systems penalize uncertainty, prompting models to give confident yet incorrect responses.
Proposed changes include redesigning benchmarks to reward uncertainty and reduce guessing.
OpenAI seeks to improve AI reliability by adjusting evaluation incentives.