Google DeepMind CEO Warns: Humanity Has a 'Precious Window' to Ensure AGI Safety

Demis Hassabis, as founder and CEO of Google DeepMind, holds a uniquely authoritative position in the AI field. His team has developed milestone AI systems including AlphaGo and AlphaFold, and he received the 2024 Nobel Prize in Chemistry for contributions to the AlphaGo project. That such a scientist at the forefront of AI publicly calls for safety reviews of AI development speaks volumes. Hassabis clearly states in his essay that AGI's arrival is no longer a distant future scenario but a 'reality that is happening.' He believes humanity is currently at a critical historical juncture — before AGI fully matures, we still have a 'precious window' to establish necessary safety frameworks and governance mechanisms.

The core proposal Hassabis puts forward is establishing a US-led international coalition specifically responsible for reviewing the safety of the most advanced AI models. This coalition's operating mechanism would be similar to the International Atomic Energy Agency (IAEA)'s regulatory model for nuclear energy: before AI models are publicly released, an independent team of experts would conduct comprehensive safety assessments to ensure models won't be misused or create unexpected risks. Hassabis believes this review mechanism should be voluntary but requires the joint participation and support of major AI labs. He particularly emphasizes that safety reviews aim not to hinder innovation but to ensure AI technology develops in directions beneficial to humanity.

Hassabis's call does not come out of nowhere but has clear real-world context. Just days before his essay was published, the Trump administration had just intervened in model releases by OpenAI and Anthropic. Both OpenAI's GPT-5.6 and Anthropic's Claude Mythos 5 were required to undergo government review before public release. This marks the first time the U.S. government has conducted pre-release review of AI models, representing a significant shift in AI regulation from 'post-hoc accountability' to 'preemptive prevention.' White House officials stated all review participation was 'voluntary,' but industry observers widely believe that under growing government attention, the line between 'voluntary' and 'mandatory' is becoming blurred.

From a technical perspective, Hassabis's concerns are well-founded. Current state-of-the-art AI models have already demonstrated astonishing capabilities: GPT-5.6's Sol model set new records in coding ability, Anthropic's Fable 5 can handle over 1 million token ultra-long contexts, and xAI's Grok 4.5 excels in agentic tool use. The rapid advancement of these capabilities means AGI may arrive earlier than expected. Once AI systems comprehensively surpass humans in cognitive abilities, traditional 'human oversight' models will no longer be effective. Therefore, establishing effective safety frameworks before AGI arrives is indeed an urgent task facing humanity.

However, Hassabis's proposal also faces numerous challenges. First is the difficulty of execution: how to ensure all major AI labs participate in reviews? How to handle safety issues discovered during reviews? What measures should be taken if a lab refuses to cooperate? Second are political challenges: government-led AI review mechanisms may spark debates about freedom of speech, freedom of innovation, and excessive government intervention. Third is the complexity of international coordination: AI is a global technology, and a single country's review mechanism cannot cover all participants. Whether other major AI powers like China and the EU will join this coalition will directly affect its effectiveness.

Notably, Hassabis's position forms an interesting contrast with his competitors. OpenAI CEO Sam Altman has long held a relatively open attitude toward AI regulation but frequently disagrees with safety researchers on specific implementation. Anthropic CEO Dario Amodei has consistently emphasized a 'Responsible Scaling Policy,' but its actual implementation effectiveness remains to be verified. xAI's Elon Musk holds the most contradictory position — he is both one of the earliest warners of AI safety risks and one of the entrepreneurs pushing the fastest AI development. In this complex landscape of interests, Hassabis's position as Google DeepMind's leader calling for unified safety standards carries a degree of neutrality and credibility.

For ordinary users and developers, AGI safety discussions may seem distant but are actually closely related to everyone's vital interests. AI model safety reviews may affect the release speed of new features, API availability, and AI service pricing. If safety reviews cause model release delays, developers may need to adjust product roadmaps. If reviews increase compliance costs, these costs may ultimately be passed on to users. Therefore, following the progress of AGI safety discussions and understanding regulatory policy trends is crucial for every AI practitioner and user. Evergreen Tools' 字数统计工具 can help analyze the content structure of policy documents, 文本对比工具 can help compare changes between different versions of policy documents.