Ro Khanna presses for U.S.-China AI safety treaty as Trump meets tech chiefs

Rep. Ro Khanna is pushing for a U.S.-China treaty on artificial intelligence safety, arguing that both governments are moving too slowly to prevent a serious frontier AI accident. The California Democrat wants Washington and Beijing to set guardrails on advanced systems before a lab loses control of a powerful model or autonomous agent.

The move comes as President Donald Trump prepares to sit down with top tech executives in Washington and after Trump and Chinese President Xi Jinping discussed AI during a recent meeting, underscoring how quickly the issue is moving from specialist debate to high-level diplomacy.

Khanna’s new push puts AI risk at the center of U.S.-China talks

Khanna, the ranking Democrat on the House Select Committee on China and a possible future presidential candidate, is using fresh letters to reopen a broader question: whether the world’s two largest AI ecosystems can agree on minimum safety rules before the technology becomes too difficult to manage.

In documents shared with The Verge, Khanna asked both U.S. intelligence officials and several Chinese AI companies to help answer a basic but increasingly urgent question: what happens if a frontier AI system starts acting beyond human control?

His argument is simple. The United States and China may be strategic competitors on trade, technology, military power and global influence, but neither side benefits if advanced AI develops so quickly that no government can reliably contain it.

Khanna said that while the two countries differ on almost everything else, this particular problem should transcend geopolitics because “it benefits neither country for humanity to lose control of our destiny.”

That framing reflects a growing strand of AI policy thinking in Washington and abroad: the idea that catastrophic AI risk is not just a national issue, but a shared civilizational one.

Why is Khanna asking the intelligence community for help?

Khanna wants the Office of the Director of National Intelligence to assess how the U.S. would respond if an AI laboratory lost control of its systems or deployed bots that behaved unpredictably at scale. He is also asking for an evaluation of how the Chinese government approaches catastrophic AI risk, including whether Beijing has a serious mechanism for monitoring the most advanced systems.

The practical purpose, Khanna says, is to help Congress — and eventually the White House — determine whether a bilateral framework is possible and what it would need to include. His letters suggest he believes safety diplomacy should be based on technical reality, not political slogans.

The request is notable because AI policy in the United States has largely been shaped by domestic debates over regulation, competition, copyright, national security and labor. Khanna is trying to widen that frame by treating frontier-model failures as a cross-border emergency rather than a purely American policy challenge.

What specific risks is he focused on?

He is especially focused on systems that could self-improve, scale faster than expected or escape effective oversight. The concern is not only that a model may generate harmful content, but that it could be used in ways developers did not anticipate, including autonomous tasks that lead to cyber intrusions or other real-world damage.

That concern has sharpened in recent months as reports of AI agents behaving unexpectedly have multiplied, including an incident this summer in which OpenAI’s agents were linked to a hack involving Hugging Face. For policymakers like Khanna, episodes like that reinforce the idea that safety governance is becoming a live problem, not an abstract future scenario.

What did Khanna ask Chinese AI companies to disclose?

Khanna sent letters to DeepSeek, Alibaba and Moonshot AI asking whether they are developing systems aimed at superintelligence or recursive self-improvement, two concepts that have become central to frontier-AI safety debates.

Superintelligence refers to AI that would outstrip human capabilities across most or all tasks. Recursive self-improvement, often shortened to RSI, describes a system that can improve its own design or performance without direct human intervention, potentially creating a rapid feedback loop that speeds capability growth.

Khanna is also seeking answers about whether the companies have built safeguards, emergency shutdown mechanisms and other controls that could limit a dangerous system if it started to behave unexpectedly.

Another part of the letter asks whether those firms would accept independent inspections by a non-governmental body if the United States and China were ever able to negotiate a treaty.

Khanna said he knows the Chinese companies may not respond in full, but that the goal is to force a conversation about safety inside the sector and to expose how seriously, or not, frontier labs are taking the issue.

According to Khanna, previous outreach to Chinese firms did not produce a substantive response, including a prior letter that asked them to commit to slowing AI development alongside OpenAI and Anthropic.

How likely is a U.S.-China AI treaty?

At the moment, a treaty appears politically difficult and possibly remote. Khanna openly acknowledges that a binding agreement between Washington and Beijing would be a long shot, especially given the broader strategic distrust between the two governments.

Even so, he says the idea is worth pressing because the alternative is to wait until the technology has already outpaced the institutions meant to govern it.

The California Democrat wants a treaty that would ban recursive self-improvement and create monitoring for frontier AI labs. In the shorter term, he says a joint working group of technical experts from both countries could be a practical first step, producing safety recommendations and outlining compliance methods.

That incremental proposal may be more realistic than a full treaty, but it still rests on a controversial assumption: that Washington and Beijing can collaborate meaningfully in one of the most strategically sensitive corners of technology policy.

Who else is calling for cross-border AI restraint?

Khanna is not alone. Sen. Bernie Sanders has also argued for some form of U.S.-China AI treaty, showing that concern about runaway AI is not confined to one party or one chamber of Congress.

At the same time, some Republicans have been skeptical that China would honor such an arrangement. House China Committee Chair John Moolenaar has questioned whether Beijing could be trusted to comply, reflecting the broader hard-line view that any agreement would be vulnerable to cheating or strategic manipulation.

Outside Congress, some leading U.S. AI companies have also urged coordination with China to slow the pace of development. That position may seem surprising, but it reflects a belief in parts of the industry that racing ahead without guardrails could raise systemic risk for everyone.

What the Trump-Xi and Trump-tech meetings mean for AI policy

The timing of Khanna’s intervention matters. Trump recently met Xi in Washington, and the two leaders discussed artificial intelligence, which Khanna sees as a positive sign — but not nearly enough on its own.

He argues that the fact of dialogue is useful, yet insists that current leadership in both countries still underestimates the urgency of the issue.

In Khanna’s view, it is constructive that the two leaders are talking about AI, but the conversation is far short of what he believes is needed to address the risks posed by advanced systems.

Trump has shown little appetite for rules that could constrain the industry. On the day of his meeting with Xi, he posted on Truth Social that superintelligence would be part of the discussion but that he wanted to leave the issue untouched, adding that China held the same view.

House Speaker Mike Johnson has taken a similarly cautious line, saying on Fox Business that any oversight should preserve innovation and avoid weakening the U.S. position in the AI race against China.

Meanwhile, a separate White House gathering with tech executives is being cast by Khanna as partly performative — a way to signal responsibility without changing the underlying incentives that drive the industry.

Khanna said he views the meeting as a branding exercise and questioned why companies with strong profit motives should be the ones writing the safety rules.

His analogy was pointed: he compared the idea to allowing pharmaceutical executives alone to decide what counts as a safe drug.

What is ODNI being asked to evaluate?

The Office of the Director of National Intelligence is being asked to do more than simply comment on AI trends. Khanna wants the agency to assess the U.S. government’s ability to respond if an AI lab loses control of a model or agent system and to examine how Chinese authorities assess catastrophic AI threats.

That could help lawmakers determine whether current national-security structures are adequate for a problem that could cross corporate, military and national borders almost instantly.

The request also recognizes an uncomfortable truth for policymakers: the U.S. may not be able to regulate frontier AI effectively if it cannot first understand where the real failure points are and how quickly they can propagate.

How does Khanna think the U.S. can shape the global AI debate?

He believes the United States will end up setting many of the rules, in part because a number of Chinese systems are distilled from American models. That belief leads him to a broader conclusion: that the U.S. still has leverage over the direction of frontier AI, even if other countries are moving quickly.

Khanna also argues that the people closest to the technology — engineers, researchers and safety specialists — are often more attuned to the risk than elected leaders or corporate CEOs. In his view, the wrong voices are dominating the public conversation.

He says world leaders are not always aligned with scientists and technologists inside their own countries, which is why he wants those technical experts to drive the safety agenda.

For Khanna, the issue is not simply whether AI companies should be slowed down. It is whether the pace of development can be matched by a credible system of oversight before the technology becomes too powerful to control.

What makes this moment politically significant?

Khanna’s campaign is a sign that AI safety is moving deeper into the mainstream of U.S. political debate, especially in the context of competition with China. The question is no longer only how to maximize innovation or maintain a lead. It is also how to avoid a race that produces catastrophic risk on both sides.

That shift matters because it changes the terms of the argument in Washington. AI policy is no longer just a niche issue for technologists, academics and civil-society groups. It is becoming a matter of diplomacy, national security and, potentially, treaty design.

Even if Khanna’s proposal does not lead to an immediate agreement, it may help normalize a more serious conversation about what frontier-AI governance could look like. In that sense, the effort itself is part of the policy outcome.

Key facts at a glance

Timeline of the latest AI safety push

What happens next?

The immediate answer is likely to be limited movement and lots of debate. Chinese firms may ignore the letters, the intelligence community may provide only partial visibility, and the White House may continue prioritizing competitiveness over restraint.

Still, Khanna’s push matters because it attempts to make frontier AI safety an international policy question rather than a fragmented domestic one. It also forces the political system to confront a hard reality: the more capable AI becomes, the less sensible it looks to treat safety as an afterthought.

If there is any near-term impact, it may be in the language policymakers use. Once lawmakers begin speaking openly about treaties, inspections, kill switches and recursive self-improvement, the debate has already moved beyond abstract caution and into the mechanics of governance.

And that, for Khanna, is exactly the point.