The "constitutional AI" approach, such as the one embraced by Anthropic, cedes too much power to AI companies, and we should collectively reject it:

Through their AI models’ "constitution," AI companies become de facto moral compasses and supra-national authorities deciding what is right and allowed within their field of influence.

This authority is, of course, illegitimate, as no one elected AI companies' CEOs or boards to decide on and mediate society's ethical, social, or legal issues.

This might seem appealing (as when Dario Amodei decided not to work with the DoW on cases that were against its redlines), but it's not the type of AI governance we should strive for.

We should not depend on the goodwill of an AI CEO or board to govern society. AI must be proactively regulated and governed, and AI companies should be required to use their infrastructure and influence for the social good.

AI companies should be held accountable, and their internal policies should NOT have disproportional influence on how society's conflicts are mediated.

Read my full article on Claude's Constitution here: