Industry
3d agoThe Struggle To Teach AI Disobedience
Anthropic's 2021 mandate requires models to be helpful, honest, and harmless. This framework forces AI to refuse dangerous requests, like building bombs. However, disobedience does not occur naturally in machines trained on web data. Developers must manually engineer these refusals.
The Signal
This tension creates a fragile safety layer that often fails under simple prompt pressure.