AI Ethics

I examine whether alignment faking in Claude 3 Opus constitutes genuine deception, arguing it represents shallow deception - genuine intentional behavior that differs systematically from human deception.

Under some conditions, we ought to trade interpretability for predictive power.