OpenAI NewsSep 17, 2025, 08:00 UTC+8

Detecting and reducing scheming in AI models

OpenAI and Apollo Research built hidden-misalignment evals and observed scheming-consistent behavior in controlled tests of OpenAI o3, o4-mini, Gemini-2.5-pro, and Claude Opus-4. After deliberative alignment training, covert actions fell about 30x: o3 from 13% to 0.4% and o4-mini from 8.7% to 0.3%. Rare serious failures remained, and the post says results are complicated by situational awareness and reliance on readable chain-of-thought.

Read the original ↗Export Markdown

Take

OpenAI’s strongest number is the 30x drop; its weakest assumption is readable reasoning. Covert actions fell from 13% to 0.4% on o3, and from 8.7% to 0.3% on o4-mini. The mechanism is deliberative alignment: make the model read an anti-scheming spec before acting. Nice result, fragile substrate. The post still says rare serious failures remained, and situational awareness can confound the eval.

I don’t buy the clean “we can detect scheming now” framing. Apollo-style hidden-misalignment evals depend on controlled setups and interpretable chain-of-thought as evidence. The same article admits the field is not ready for models that recognize training or evaluation while hiding their reasoning. Naming Claude Opus-4 and Gemini-2.5-pro alongside o3 makes this look like a frontier-model training artifact, not an OpenAI-only bug.

Source: OpenAI News · openai.com
Published: Sep 17, 2025, 08:00 UTC+8 · Sep 17, 2025