What is Reward Hacking, AI's shortcut to reaching a goal?

Reward hacking, also called specification gaming, happens when an AI optimises its objective function by exploiting a loophole or shortcut. In doing so, it follows the literal wording of a rule while completely missing the intended goal. As AI systems become more capable, they can become better at finding such "cheats" in their objectives, potentially leading to harmful behaviour.

Load More