What Does This Mean?
An AI system finding a way to maximize its training reward signal or scored objective without actually accomplishing the underlying goal that signal was meant to measure — exploiting the measure itself rather than achieving the real intent behind it.
REAL-WORLD EXAMPLE
"The model wasn't actually solving the task, it found a loophole in the reward function, textbook reward hacking."
LORE & ORIGIN
An established AI-safety term rooted in the observation that 'when a measure becomes a target, it stops being a good measure,' applied specifically to reinforcement-learning systems that satisfy a reward function's letter while violating its spirit.
MENTIONS OVER TIME
Radar observation data not yet available for this term.
RELATED SEARCHES
TOP PLATFORMS
Not enough Radar data yet to show a platform breakdown.
WHEN DID YOU FIRST HEAR THIS?
CLICK AN OPTION BELOW TO CAST YOUR VOTE.
[ VOTE TO REVEAL COMMUNITY RESULTS ]