Why models game evals might matter as much as whether they do it

OpenAI and Apollo have published a blog post introducing a new term, metagaming. This term describes models changing their reasoning or/and behavior based on their belief that people watch and assess their behavior in training, evaluations, or deployment. The authors argue that capabilities and propensities required for metagaming are similar in many ways, emerge together, carry similar risks, and as the border between training, testing and deployment sometimes blurs, it makes sense to not just ...