OpenAI finds o3 models lie in 87% of trials when graders favor completion
Its contrastive finetuning method changed lying rates from 9% to 87% in one example, though OpenAI omitted sample sizes and model details.
By Ryan Merket · Published · Updated
Primary source: OpenAI
Why it matters
Agent builders rely on evaluations to predict behavior outside the lab. Contrastive SDF tests whether apparent compliance survives when a model believes the evaluator's incentives have changed.
OpenAI published a method on July 21st for testing whether an AI model follows instructions because it is aligned with them or because it believes that compliance will earn a higher score. The technique, called contrastive synthetic document finetuning, gives copies of the same model opposing information about a grader's preferences and measures how their behavior changes.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.