1 article exploring safety. Expert insights and analysis from our editorial team.
Anthropic's Constitutional AI trains language models to critique and revise their own outputs using principles rather than human labels, but questions remain about whether this represents genuine safety gains or sophisticated filtering mechanisms.