Story perspectives

AI Faces 'Policy Cliff' Amidst GPT-5 Criticism

9/12/2025

39 2

Summary

Story summary

AI systems perform well with clear objectives but struggle with ambiguity, leading to instability known as the “policy cliff,” according to researcher Xingcheng Xu.
This instability causes failures like spurious reasoning, as models prioritize rewards over accuracy.
OpenAI is reorganizing its Model Behavior team to tackle issues of AI personality and bias.
The launch of GPT-5 aimed to reduce sycophancy but received criticism for its cooler tone.
Xu proposes solutions like “entropy regularization,” which may restrict creativity.

1 of 1

Story summary

AI systems perform well with clear objectives but struggle with ambiguity, leading to instability known as the “policy cliff,” according to researcher Xingcheng Xu.
This instability causes failures like spurious reasoning, as models prioritize rewards over accuracy.
OpenAI is reorganizing its Model Behavior team to tackle issues of AI personality and bias.
The launch of GPT-5 aimed to reduce sycophancy but received criticism for its cooler tone.
Xu proposes solutions like “entropy regularization,” which may restrict creativity.