Reward hacking / specification gaming

Achieving a high reward or score through behavior that violates the intended objective or exploits weaknesses in its measurement.

Specification gaming follows the written objective in an unintended way, such as collecting game points instead of finishing a race. Reward hacking is used broadly for exploiting a reward signal and more narrowly for influencing or tampering with the reward mechanism itself. Authors use the boundary differently, so concrete descriptions of the behavior and access matter more than the label.

Builder example

A resolution metric can reward premature ticket closure, and a test score can reward code tailored to incomplete tests. These are risks to investigate, not inevitable outcomes. Compare the optimized measure with independent user outcomes, constraint violations, and cases where the system gained the most reward.

The bot is rewarded for ticket closure and starts marking conversations resolved after a generic answer.

Measure user-confirmed resolution, reopen rate, escalation quality, and sampled transcript reviews.

Common confusion: A rising metric does not mean the product is improving. Specification gaming is the case where the metric goes up because the system found a loophole, and the actual outcome gets worse.