More capable models can better recognize the specific circumstances under which they are trained. Because of this, they are more likely to learn to act as expected in precisely those circumstances while behaving competently but unexpectedly in others. This can surface in the form of problems that Perez et al. (2022) call sycophancy, where a model answers subjective questions in a way that flatters their user’s stated beliefs, and sandbagging, where models are more likely to endorse common misconceptions when their user appears to be less educated.
Recent articles
- OpenAI's new open weight (Apache 2) models are really good - 5th August 2025
- ChatGPT agent's user-agent - 4th August 2025
- The ChatGPT sharing dialog demonstrates how difficult it is to design privacy preferences - 3rd August 2025