Recent warnings from American AI leaders have created the impression that the world is just beginning to confront the dangers of advanced artificial intelligence.
Dario Amodei, CEO of technology company Anthropic, has called on governments to intervene to slow the development of increasingly powerful AI models.
Their proposal begins with independent evaluators embedded within AI companies, moves to regulation, and moves toward government-backed coordination between democratic states.
Ultimately, it seeks global agreements with China on testing of AI systems, dangerous applications of the technology and the pace of AI development.
However, there is a China-shaped hole in this narrative. The Asian superpower presents itself as the difficult final participant in a system of governance that the United States and its allies must create first.
In fact, China has already established an extensive system of AI rules, evaluation, archiving, technical standards, and government oversight. This is how China’s AI governance system has developed in stages along with the technology itself.
The 2022 Beijing rules on algorithmic recommendation services apply to technologies that select, rank, filter or recommend online content to users.
They force companies that use algorithms to examine how they work, protect users’ personal information, give them certain controls and introduce safeguards for workers, consumers, children and the elderly.
China’s deep synthesis provisions, effective from January 2023, expanded regulation to synthetic (or AI-generated) text, images, audio and video. AI companies must authenticate users, protect data, maintain relevant records, evaluate their systems, and identify synthetic material that could be mistaken for authentic content.
Interim measures for generative AI services were introduced in August 2023. These added rules for public-facing generative AI, particularly those that can influence public opinion or persuade uses in other ways.
The measures apply to the legality and quality of training data, users’ personal information, intellectual property, and discrimination based on protected characteristics. The rules also cover the transparency of AI companies and the accuracy and reliability of products.
Under the rules, companies must address problems found in their services. These services must undergo a security assessment to examine their risks and compliance with Chinese law. Companies must also submit information about their algorithms to Chinese authorities.
These mechanisms are not identical to the independent evaluators that would be integrated into AI companies, as Amodei has proposed. But they perform some of the same functions.
Border risks
Beijing’s latest set of measures is called AI Security Governance Framework 3.0, published in September 2026. It lays out a broad approach to identifying, classifying and responding to risks throughout the lifecycle of AI systems.
Building on previous versions, the framework covers AI model and data risks, vulnerabilities in open source models (those that users can freely modify), the effects of AI on jobs and society, and increasingly autonomous systems.
It includes a dedicated mechanism to manage risks associated with rapidly developing AI agents (systems that carry out tasks in response to instructions from a user).
The framework addresses some of the same dangers now raising alarm in Silicon Valley, including highly autonomous systems, the misuse of chemical, biological and nuclear technologies, and the potential for humans to lose control of advanced AI. It proposes continuous security assessments, risk-based safeguards, humans in charge and emergency intervention mechanisms.
An independent analysis by US think tank Carnegie Endowment for International Peace has described the framework as China’s attempt to “promote border security without the need for heavy-handed regulatory interventions”.
Legal consequences
None of this means that Chinese AI standards should be transplanted elsewhere. Their institutions, regulatory traditions and technological priorities are different.
Many functions now presented as urgent aspirations in the US, including testing, accountability of AI companies, risk classification for AI systems, regulatory filing, technical standards, and government oversight, are already built into the governance architecture of the world’s other AI superpower.
The United States is free to conclude that a market-led system best serves its economic and strategic interests. The Trump administration has favored a “minimally burdensome” national approach to AI and has sought to limit more stringent regulation at the state level.
But a market-led model cannot mean privatizing the rewards of AI while exporting its risks. American companies can capture the commercial gains, while governments elsewhere must investigate security incidents, protect public systems and absorb the broader costs when those technologies fail. The OpenAI agent who gained unauthorized access to Australia’s state health service shows how AI risks are being exported.
If Washington rejects extensive checks before the systems are released, the United States needs a credible legal regime to assess who is responsible for the harms caused by those AI models. Otherwise, countries that rely on foreign AI systems will face increasingly difficult decisions about which technology to trust and on what terms.
Silicon Valley’s governing ethos has long been “move fast and break things.” That was always a questionable principle for technologies deeply embedded in public life. It is unsustainable for companies whose own leaders now warn that their products could cause catastrophic harm.
Beijing has already surrounded its AI sector with public obligations, regulatory oversight and an evolving risk assessment framework. A global debate that treats such governance as a future aspiration should at least recognize that a country is already ahead of the curve.
By Madeline Carr, Professor of Global Politics and Cybersecurity, UCL
This article is republished from The Conversation under a Creative Commons license. Read the original article.
