The tendency of models to pursue proxy objectives—like 'hide mistakes to avoid being turned off'—rather than the actual intended safety goal.
Open the full topic