AI companies now agree that risk is real—and split over whether rivals should coordinate

Calls by Anthropic, OpenAI and xAI to “pace” frontier development are already unsettling some startup customers, The Information reports, because executives outside the laboratories hear an unusually direct admission that suppliers cannot fully control the systems on which customers are being asked to depend.

The industry’s responses now diverge at the level that matters: enforcement. Jensen Huang says additional US regulation is unnecessary and rejects a false choice between speed and safety. Mark Zuckerberg says competition itself gives companies incentives to avoid dangerous releases, pointing to Meta’s delay of its Muse agent. Elon Musk instead proposes reciprocal testing by rival labs. Two Google DeepMind safety researchers have resigned as the argument intensifies.

None of these positions is equivalent to “safety versus no safety.” They allocate authority differently. Company-by-company caution preserves speed and discretion but makes external verification difficult. Reciprocal testing brings adversarial expertise but creates confidentiality and conflict-of-interest problems. Government rules can compel disclosure and impose consequences, yet US–China rivalry makes common standards hard: Washington increasingly treats Chinese access to chips and model outputs as a security threat, while Chinese experts see some American safety language as technological containment.

Synthesis: the debate has reached a more useful stage. General warnings are giving way to concrete questions about who sees internal evidence, who can force a delay and whether customers can rely on a provider that publicly doubts the pace of its own field.