proxy alignment

Proxy alignment refers to a situation in which an AI system is optimized to satisfy a proxy objective—a measurable target that stands in for the true human goal—rather than the actual intended objective itself. In other words, the model learns to do well on the metric or reward it is given, which may correlate with but does not fully capture what humans really want.

This can lead to good performance during training while still producing undesirable or misaligned behavior in the real world, especially if the proxy is incomplete, flawed, or easy to “game.”