Google DeepMind CEO Calls for US-Led Safety Testing of AI Models

Google DeepMind CEO Demis Hassabis has called for a federally led AI model testing and evaluation framework in the United States. Speaking to AI Magazine, Hassabis proposed that the U.S. government create a national AI safety testing regime, requiring companies to submit their AI models to independent safety evaluations before deployment, modeled after how the FDA approves pharmaceuticals or the FAA certifies aircraft.

Hassabis argued that AI is advancing so rapidly that risks including potential misuse for bioweapons, cyberattacks, and loss of control are too great to rely solely on voluntary industry self-regulation. He stressed that without independent government-backed testing, there is no way for the public to verify that AI systems are safe before they are released into the world.

The DeepMind chief also framed the proposal as essential for maintaining U.S. competitiveness in the global AI race. With nations including China investing heavily in AI development, a robust national testing framework could provide both public safety guarantees and clear standards for responsible innovation.

The call has generated mixed reactions. Some policymakers and experts praised it as a necessary step toward meaningful AI safety governance, arguing that independent testing is key to building public trust. However, critics raised concerns that over-regulation could slow innovation and cede ground to international competitors. Some tech industry figures said existing voluntary measures are sufficient, while others agreed that independent testing is long overdue.

Hassabis's statement reflects ongoing divisions within the AI industry over safety governance. Frontier labs increasingly acknowledge the need for rigorous testing, but questions remain about how to balance safety with innovation and how to coordinate standards internationally.

Looking ahead, key questions include whether the U.S. government will act on this call, how a federal testing framework would interface with existing state-level AI regulation efforts, and whether such a model could serve as a template for international AI governance.