The industry now shares a warning, but not a binding plan.
SAN FRANCISCO, CALIFORNIA
Dario Amodei, Demis Hassabis, Sam Altman and Elon Musk have publicly acknowledged the need for greater caution as artificial-intelligence systems become more autonomous and capable. Their statements bring together the leaders of Anthropic, Google DeepMind, OpenAI and xAI, but the apparent consensus conceals substantial differences over who should supervise development and when intervention should occur.
Anthropic CEO Dario Amodei offered the most concrete internal commitment. His proposal would give independent evaluators, such as the nonprofit research organisation METR, employee-level access to models and company risk teams. Evaluators would work inside Anthropic but retain the ability to publish findings without corporate editorial control. Broader restrictions, however, would depend on cooperation among competing laboratories and government authorization.
Google DeepMind CEO Demis Hassabis favours an external institution. His plan calls for a federally supervised, industry-financed standards body modelled partly on the US financial regulator FINRA. Developers would submit frontier models for testing up to 30 days before release, with evaluations covering cybersecurity, biological risks, deception and attempts to bypass safeguards. Participation would begin voluntarily and could later become mandatory for models sold in the United States.
OpenAI CEO Sam Altman endorsed Amodei’s proposal and said OpenAI would also provide independent evaluators with employee-level access, although operational details remain forthcoming. Elon Musk’s response was considerably less specific: the xAI founder said Amodei was right without identifying which recommendations he supported or announcing any corresponding commitment.
The four leaders broadly agree that advanced systems require scrutiny beyond the companies building them. Yet none has committed to an immediate reduction in development speed. Anthropic conditions meaningful restraint on coordination, DeepMind assigns the decision to an institution that does not yet exist, OpenAI promises further measures and xAI has offered only general approval. The emerging consensus therefore marks an important rhetorical shift, but its credibility will depend on enforceable standards, independent authority and measurable action.
Responsibility begins when warnings become commitments.