DeepMind Proposes Global AI Watchdog to Review Models Before Release

DeepMind CEO Demis Hassabis unveiled a proposal for an international watchdog organization designed to review and test the most advanced AI models before they enter the public domain. The watchdog would be staffed by independent technical experts, financed by leading AI companies, and empowered to spend up to 30 days examining each new model. The proposal quickly attracted support from major players in the industry: OpenAI, led by Sam Altman, described it as thoughtful, while Elon Musk called it a good starting point for wider discussions. Hassabis is scheduled to meet with US policymakers in Washington the following week to discuss turning the framework into a functioning oversight system.

Hassabis explained that he chose to publish the plan in response to mounting concerns about Anthropic's Mythos model, which possesses advanced cybersecurity capabilities. He characterized Mythos as a warning for society and warned that increasingly powerful AI systems could create biosecurity risks within the next couple of years. Although Google DeepMind has already been in dialogue with government officials and AI security institutes regarding its newest models as they achieve certain scores and benchmarks, Hassabis argued that continuing to handle releases through an informal, case-by-case process would not remain viable. Under his proposal, the watchdog could introduce stronger safeguards if risks escalate, potentially including coordination among leading AI laboratories to slow development.

The path forward faces significant hurdles. Anthropic and OpenAI have already slowed the broad release of their latest models under pressure from the Trump administration, and the European Union and US have begun discussions over access to American AI technology. Wider adoption of the watchdog framework remains uncertain because the US, EU, and China have pursued divergent regulatory strategies, and Congress has not yet enacted comprehensive federal legislation governing AI. Hassabis acknowledged that the publication of the proposal represents only the initial step, recognizing that translating the framework into a formal, operational oversight structure would require political alignment and ongoing negotiation among governments and industry leaders.

DeepMind's watchdog proposal reflects growing tension between the pace of AI development and the perceived need for oversight. Hassabis framed the idea as a response to the limitations of ad hoc governance: while Google DeepMind has been discussing its newest models with government officials and AI security institutes as systems reach certain scores and benchmarks, he argued this approach may not scale as capabilities advance. The model's timing is significant—it comes as Anthropic and OpenAI have already delayed broad releases of their latest models under pressure from the Trump administration, signaling that government scrutiny is already reshaping release strategies.

However, the proposal must navigate fragmented global regulation. The US, EU, and China have taken different approaches to AI governance, and Congress has yet to pass meaningful federal legislation on the technology. These jurisdictional differences make the creation of a truly international watchdog body difficult, though Hassabis expressed belief that sufficient momentum may now exist to establish a formal structure. For AI developers, the proposal underscores that testing requirements and government scrutiny are becoming material business considerations; for policymakers, it offers a concrete model for coordinating oversight without centralizing power in any single government.