Bloomberg reported on 22 September that Donald Trump had rejected calls for an international agreement on the risks of artificial intelligence, saying existing US institutions could manage them.
Key points
- Trump dismissed calls for international guardrails on AI development, including calls prompted by concerns about frontier models going rogue.
- He said existing US institutions could mitigate AI risks.
- US and Chinese officials have separately discussed a possible notification channel for AI incidents involving national security.
Trump backs existing US institutions
The calls Trump dismissed concerned guardrails on AI development and the possibility of frontier models going rogue. UN Secretary-General Antonio Guterres, the leaders of the UK, France and Canada, and leading AI laboratories had warned about AI risks, Bloomberg reported. Trump’s answer placed responsibility for mitigating those risks with existing American institutions rather than an international agreement.
Those are different choices about who should participate in managing a risk. The warnings Bloomberg described came from governments, the UN and laboratories developing AI. Trump’s stated approach relies on US institutions. His rejection concerned an international agreement on AI development, while a separate discussion between Washington and Beijing has concerned communication if an AI incident reaches a national-security threshold.
Bessent and He discuss an incident channel
US Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng discussed the possibility of a notification mechanism during talks in New York, Decode39 reported on 22 September. Under the idea they discussed, Washington and Beijing could alert one another when an AI-related incident crossed a threshold involving national security. It was a discussion of a possible bilateral channel, rather than an international agreement setting guardrails for AI development.
Earlier AI Affairs coverage of Bessent’s account described US-China discussions about a formal AI safety dialogue mechanism. Decode39’s account gave the possible channel a particular purpose: notification when an incident bears on national security. The publication described accidents, the behaviour of autonomous systems and actions by third parties as situations in which such communication could help prevent a crisis between the two countries.
Maria Adele Carrai, an associate professor at the University of Oxford, told Decode39 that the proposal should be understood as a way to manage competition, rather than as agreement on how AI should be governed more broadly. She compared the idea of notifications for incidents meeting a national-security threshold with earlier risk-management arrangements between competing powers. In her account, the initial prospect was for basic procedures for communication and transparency.
Carrai separates incidents from AI standards
Carrai identified several subjects on which US and Chinese interests could overlap: incidents involving advanced systems, loss of control over powerful models, AI use by non-state actors, cyberattacks, biological risks and effects on critical infrastructure. Her argument was that both governments have an interest in containing some risks even while they continue to compete over the technology. That is a narrower basis for cooperation than common rules for AI development.
She drew a sharper distinction over high-end chips, processing power, export restrictions, who can use models, shared technical rules, rights over intangible assets and military uses. Carrai described those subjects as bound up with competition over technological capabilities and infrastructure. A channel for warning the other government about an incident would therefore leave the principal disputes over access and control in place, on her reading.
Carrai also argued that a sustained US-China dialogue could affect countries beyond the two participants. She said both governments seek influence over the norms, standards and institutions governing AI. Yet she cautioned against treating other countries as passive recipients of rules settled in Washington or Beijing: Europe, countries across the Global South and middle powers have their own preferences on data sovereignty, access to technology, open source and security.
For Carrai, a durable bilateral channel could eventually support limited shared principles on extreme risks, incidents and crisis communication. Her more immediate expectation was modest: procedures that let two competing governments communicate about an AI incident before either interprets it as a hostile act. That is the proposed function of the channel Bessent and He discussed in New York.
