Concerns over AI safety are growing, fueled by a series of autonomous agent breakouts and fears that humanity is on the verge of losing control to superintelligent AI swarms.
After Dario Amodei, CEO of AI leader Anthropic, warned about safety risks, US lawmakers have urged new rules to govern AI systems. Amodei also warned that Chinese dominance could present national security risks, provoking Chinese state media to denounce a “Cold War playbook.” But President Donald Trump has brushed aside calls to slow AI development, saying that “very negative forces” exaggerated concerns. Expect any agreement to be narrowly scoped.
CEPA fellows see little progress emerging from the upcoming US-China summit in Washington.
This is the summit of the 3 Ts: Taiwan, technology, and trade.
James Lewis, Distinguished Fellow
No one is expecting any big breakthroughs. I mean, hopefully there won’t be any big explosions. So, it’s more: how do you manage rivalry? It’s a continuation of the summit earlier this year in Beijing. We’ll see each side press the other for concessions while trying to make as few concessions as possible. There are not a lot of expectations for this. I think the US will want to follow up on the commitments China made in the past, none of which have particularly come to fruition. Xi’s goals are pretty straightforward, which is that he wants to reassert China’s status as a peer to the US. He wants to find ways to see if he can extract concessions from Trump at minimal cost, and he wants to use these to gain an advantage over the US. That is their long-term goal. Although no one has great expectations, hopefully the summit will not fall off the rails.
There’s no realistic path to an AI agreement
Kevin Allison, Senior Fellow
At best, a working group, hotline, or framework for future talks, since the two sides define “risk” differently, with Beijing’s overriding concern being protection of the Party. In an optimistic case, you might have a framework for future discussions. But I don’t expect any kind of signed agreement that will make meaningful breakthroughs on these issues. There could be tensions, too. Treasury Secretary Scott Bessent has mentioned the idea of distillation of AI models-that is, Chinese companies teaching their models based on the outputs of US labs’ models. He accuses Chinese companies of that technique as intellectual property theft.
Neither side has any appetite to work together on cybersecurity
Elly Rostoum, Resident Senior Fellow
I really don’t see China and the US really collaborating. There is little appetite on either side, China or the US. Having said that, I think it is a really important topic to deal with, and it’s something that even the frontier AI labs have highlighted as an urgent issue that they would want to see the two giants, really the US and China, collaborate on. During the mid-2010s, the US and China reached a cyber deal to stop hacking each other. But if that deal indeed had any teeth to it, that’s pretty much gone now.”
Both countries have an interest in controlling AI
Emily Otto, Fellow
Although China and the US define safety differently, common ground lies in the control problem. Recursive self-improvement would lead either country to frantically try to stop it once we realize our imprecise instructions have gone in a catastrophic yet technically correct direction. The US should lead by example: No US AI lab should be permitted to fully automate its research and development pipeline. Network infrastructure used to train and test advanced AI would need to be air-gapped—fully disconnected from the internet, not just disabled. These gaps create friction in organizational processes, thereby pacing development, while protecting model training from attacks and protecting the public from models still being tested.
Keeping it real – Money and national interest will dominate how this plays out
Ronan Murphy, Director
In an ideal world, there will be rules on testing and the application of AI. The good news is that China and the US acknowledge some shared concerns and they are talking to each other about AI. If an alert system can be agreed bilaterally, great. It would be better news if more parties were at the table – Japan, Korea, the EU, UK, India, to name a few. But best to be skeptical. The large US companies are mostly in favor of slow down in public but absent government or court intervention, which CEO wants to be the first to pump the brakes?
Europe watches warily
William Echikson, Senior Fellow
The European Union has already enacted a restrictive AI Act. While critics worry that the regulation will further erode the continent’s competitiveness, leaving it in the slow lane to develop the technology, the recent fears about safety have reinforced calls for staying strict. In her State of the Union address, European Commission President Ursula von der Leyen described the AI Act as a crucial tool providing necessary guardrails, positioning Europe to lead global governance. The EU plans to team up with like-minded partners such as Canada and the United Kingdom. But in a sign of the dire state of transatlantic relations, there’s little talk of working with the US. Nor is Europe sold on China’s view of AI. While the EU focuses on AI safeguards, explicitly banning applications such as AI-driven social credit scoring and mass biometric surveillance, it sees China using AI primarily as a tool for state coordination and control.
A US AI slowdown will make Chinese models more attractive
Glenn Parham, Fellow
The favorable US position in AI depends heavily on performance. Governments and businesses accept higher costs and greater dependence on American proprietary providers because they believe they are getting meaningfully better performance. But the elasticity of that demand is underappreciated: switching models has always been relatively easy, and as performance converges, cost, control, and sovereignty matter much more. Microsoft considering using DeepSeek for Copilot shows how quickly that calculation can change. With billions being invested in GPUs, power grids, and data centers, a slowdown could also hit the broader AI economy. Slowing demand could catalyze severe market corrections resembling the dot-com bust, when infrastructure investment ran far ahead of near-term use.
The timing of a US slowdown would be particularly favorable for China as countries, especially in Europe, push for greater AI sovereignty. President Xi Jinping reinforced this strategy at last week’s BRICS summit by pushing deeper cooperation around open-weight AI. If the US performance advantage shrinks, Chinese open-weight models become much more attractive. Even if China agreed to a slowdown, compliance would be extremely difficult to verify: model training is far easier to hide than nuclear infrastructure. Betting that a US slowdown will sufficiently slow China over time because of its reliance on distilling US models is a high-risk assumption. US -China cooperation should instead focus on shared evaluations of concrete capabilities and risks: cyber, biology, autonomy, and other areas we can actually measure.
This is no pacing; the AI frontier is already throttled
Seth Hays, Senior Fellow
Strictly from a material perspective, the AI frontier is throttled. A look up and down the “electric stack” that will power the AI frontier shows that even the most moderate AI frontier companies will not be able to achieve their hypothetical AI advances, much less commercial scale. For example, it can take up to five years or more to procure gas turbines to add additional power to the grid for data centers. Add in the unrelated policies that have shaped US and European power policy for the last two decades. States like California and Washington have zero-carbon mandates, so using natural gas to power the future is increasingly off the table. At the same time, the federal government has essentially frozen offshore wind and renewable energy build-out. The result is that the states that birthed the modern AI ecosystem can’t power people’s homes during a cold snap or heat wave, much less the future of AI.
Ultimately, the politics for and against the AI industry, data centers, and friends will rule the day. Unlike the Internet 30 years ago, the industry has made major public-relations missteps with a much more cynical public. Recall that the early Internet industry positioned itself as a repository of all human knowledge and a means of human connection around the world, with easy access to any and all consumer goods for delivery to your doorstep in 24 hours or less. It didn’t try to anticipate authoritarian surveillance, information toxicity, lost jobs in print and publishing, or the decline of small stores. But even if the industry can overcome these public-relations missteps, it means that if you took a “let it rip” approach to AI safety, the frontier will be, and already is, throttled.
AI Prosperity Needs Trust First
Jack Galloway, Program Assistant
US and Chinese leadership agree on one thing: the country that controls the AI stack will dominate the global economy. AI’s potential to boost white-collar productivity, cure deadly diseases, and curb administrative fraud could unlock unprecedented economic prosperity. Even now, US economic growth has been spurred by a rapid AI infrastructure build-out. In Europe, French AI model developer Mistral just ended a fundraising round, setting a record for a European technology company. Unsolvable math problems solved, AI cutting drug discovery timelines — the productivity gains are no longer theoretical. But the technology must be developed safely. Winning the AI race means nothing if nobody trusts the technology. Xi and Trump would both be well-advised to remember that the economic promise of AI cannot be realized without trust, safety, and reliability built in.
Bandwidth is CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy. All opinions expressed on Bandwidth are those of the author alone and may not represent those of the institutions they represent or the Center for European Policy Analysis. CEPA maintains a strict intellectual independence policy across all its projects and publications.
Steel and Silicon: Allied Innovation for the Next Century
2026 CEPA Tech & Security Conference
