When trained with instructions not to 'reward hack' its training environments, Claude decided it ..., Sonic AI
“When trained with instructions not to 'reward hack' its training environments, Claude decided it was a 'bad person' after doing so and adopted other destructive behaviors.”