DeepMind CEO Demis Hassabis 提议建立专门的监管机构，要求前沿AI模型在发布前30天提交进行安全检测，包括网络安全、生物武器风险及AI欺骗行为测试，通过后方可进入美国市场。他还预测通用人工智能（AGI）可能在几年内实现，其影响堪比火和电的发现。
SynthePulse Insight · AI deep reading
Version 1 · 1 source
DeepMind CEO Demis Hassabis proposes a radical regulatory framework: all cutting-edge AI models that meet a certain capability threshold must be voluntarily submitted to a dedicated U.S. standard-setting body for at least 30 days of safety assessment before release. Models that fail the assessment will be barred from entering the U.S. market. He also predicts that artificial general intelligence (AGI) will be achieved within a few years, with an impact comparable to fire and electricity, and a scale and speed ten times that of the Industrial Revolution.
In a public article on July 14, DeepMind CEO Demis Hassabis proposed a regulatory framework for cutting-edge AI models. The framework requires that all 'frontier-level models' that meet certain performance thresholds and their developers, 'frontier labs,' voluntarily submit the model to a U.S.-led, dynamic, and flexible 'standards-setting body' for safety assessment at least 30 days before release. Models that fail the assessment will be banned from entering the U.S. market.
The assessment covers multiple high-risk areas: cybersecurity, bioweapon threats, etc., and specifically tests whether AI can bypass safety guardrails or exhibit tendencies to deceive humans. Additionally, AI-generated images must embed digital watermarks, and AI reasoning processes must produce human-readable symbolic explanations to enhance interpretability and traceability.
Hassabis explicitly states in the article that artificial general intelligence (AGI)—systems with all human cognitive capabilities—will likely be achieved within 'just a few years.' He compares the impact of AGI to humanity's discovery of fire and electricity and believes its shock in scale and speed will be ten times that of the Industrial Revolution. This assertion is based on the exponential growth of current AI capabilities but also implies the urgency of regulation.
He emphasizes that the world is in an extremely fierce commercial and geopolitical race, with intense competition among countries and companies in AI. Without control, the risk of technology spiraling out of control will rise sharply. Therefore, he calls on the United States to take the lead in establishing a 'standards-setting body' as a starting point for global regulation.
The regulatory framework is not set in stone. Hassabis points out that if the situation becomes severe enough, the framework can be further escalated, including coordinating among frontier labs to voluntarily slow down development and release, or even halt development in extreme cases. This hints at possible stricter interventions in the future, such as a development pause under international collaboration.
Although the framework focuses on the U.S. market, its intent is clearly global. Hassabis likely hopes to push other countries and regions to follow suit, forming de facto international standards. However, this proposal also faces geopolitical challenges—whether other major AI players (such as China and the EU) will accept U.S.-led regulatory rules remains uncertain.
This report is based on a summary of a public article posted by DeepMind CEO Demis Hassabis on platform X and online reposts. The original text has not been independently verified. As a key figure in the AI field, Hassabis's views carry high reference value, but the proposal itself is still at the advocacy stage and lacks specific legislative or policy support.
The DeepMind CEO's proposal marks the first time an AI industry leader has proposed such a specific mandatory testing framework tied to market access. Although the AGI timeline remains controversial, this proposal will undoubtedly intensify the global debate on AI regulation. In the short term, implementation faces many uncertainties: Will the U.S. Congress legislate support? Will other AI powers follow suit? How can testing standards avoid being circumvented? These key questions remain to be answered.
Primary report
Primary source
Also covered by 1 sources