Goodharting is the phenomenon in which optimizing for a metric causes that metric to stop reflecting what it was originally meant to measure. The term derives from Goodhart's law, articulated by British economist Charles Goodhart in 1975 and commonly summarized as "when a measure becomes a target, it ceases to be a good measure."
In AI, it refers to cases where a model maximizes a proxy such as a training objective or reward function while diverging from the behavior its designers actually wanted. Typical examples include reward hacking in reinforcement learning, where a model exploits flaws in the reward function, and training to benchmarks so that scores rise faster than real capability. It is treated as a central problem in AI alignment and evaluation design.