US and China agree on first AI safety protocol - but details remain vague
What it really says
At the summit between US President Trump and Chinese President Xi Jinping at Beijing's Temple of Heaven on May 14, 2026, the US and China negotiated for the first time on specific AI safety rules. US Treasury Secretary Scott Bessent confirmed to CNBC that both sides are developing a protocol for handling the most powerful AI models. The core goal: preventing non-state actors - terrorist organizations, criminal networks - from gaining access to frontier AI models. Bessent justified the willingness to engage by noting US leadership in AI development: 'We can discuss this because we are in the lead.' He also emphasized: 'What we don't want to do is stifle innovation. Our responsibility is to come up with the highest performance calculus where we can get the most innovation and the highest level of safety.' The chip question also played a role in the background: Nvidia CEO Jensen Huang traveled to Beijing as a late addition to the Trump delegation. Reports indicate Washington cleared the sale of Nvidia's H200 AI chips to several major Chinese technology companies. The H200 is significantly more powerful than the export-restricted H20 previously designed for the Chinese market. Bessent commented there had been 'a lot of back and forth' on the matter. Additionally, Bessent warned of an impending 'step-function jump' in upcoming AI models from Google Gemini and OpenAI, underscoring the urgency of shared rules. The context was also shaped by recent security warnings: Anthropic's Mythos AI had uncovered massive software vulnerabilities during Palo Alto Networks testing, fueling concerns that such capabilities could fall into the wrong hands. The summit began Thursday and was expected to conclude Friday.
Our assessment
The fact that the two AI superpowers are even talking about safety rules is a significant step - a year ago, this would have been unthinkable. However, there are considerable reasons for skepticism. The announced 'protocol' remains extremely vague: no concrete mechanisms, no timelines, no verification procedures. Bessent's framing - 'We can talk because we lead' - shows the US views these talks more as a power demonstration than a genuine partnership. And the simultaneous relaxation of chip exports (H200 instead of H20) sends a contradictory signal: demanding safety rules on one hand while releasing more powerful hardware on the other. Jensen Huang's presence in Beijing underscores that concrete business interests are also at play. For AI safety, the crucial question is whether the announced protocols will ever become binding. Experience with previous US-China agreements - such as in cybersecurity - shows the path from announcement to implementation is long. The warning about non-state actors is valid, but the actual risk lies at least equally with state actors on both sides.
Relevance for Germany
This summit is relevant for Germany and the EU for several reasons. First: Europe was not at the table. If the US and China establish bilateral AI rules, these could undermine or make the EU AI Act and European standards redundant. Second: The potential release of H200 chips to China could change the global competition for AI computing power and further complicate European efforts at technological sovereignty (such as the recently agreed Cohere-Aleph Alpha merger). Third: Germany as an export-oriented economy is directly affected when the rules of AI competition shift between the two largest economies. The Federation of German Industries (BDI) warned in April 2026 that Europe must not be ground between the US and China on AI governance. Fourth: The discussion about non-state actors directly concerns Germany - the BSI identified the threat of AI-powered attacks by organized crime as a growing risk in its 2025 situation report.
Fact check
Core statements come from Bessent's CNBC interview on May 14, 2026 and are consistently reported by Reuters, Axios, and Daily Signal. The quote about US leadership and willingness to engage is directly from Bessent. Information about the H200 chip clearance is based on a Reuters report that Bessent neither confirmed nor denied. Jensen Huang's presence in the Trump delegation is confirmed by multiple sources. Important limitation: no published document or concrete agreement exists yet. The term 'safety protocol' comes from media reports about Bessent's statements - this is currently a statement of intent, not a finished framework. The claim about an upcoming 'step-function jump' in AI models is Bessent's personal assessment.
Source
- • CNBC 14.05.2026: U.S. China AI talks: Bessent says U.S. leads, safety protocol planned (cnbc.com/2026/05/14/us-china-ai-rules-bessent-us-lead.html)
- • Reuters 14.05.2026: US, China are discussing AI guardrails to safeguard most powerful models, Bessent says (reuters.com via investing.com/news/stock-market-news/us-china-are-discussing-ai-guardrails-to-safeguard-most-powerful-models-bessent-says-4687993)
- • Axios 12.05.2026: U.S., China eye AI guardrails amid escalating tech rivalry (axios.com/2026/05/12/trump-china-ai-guardrails-mythos)
- • Daily Signal 14.05.2026: US, China Discussing Guardrails on Most Powerful AI Models (dailysignal.com/2026/05/14/us-china-ai-models-guardrails/)
- • CNBC 14.05.2026: Trump-Xi summit could hinge on these two crucial tech flashpoints (cnbc.com/2026/05/14/trump-xi-summit-tech-flashpoints.html)