The United States and China are negotiating a direct communication channel for AI safety incidents. Think of it as a hotline for when things go wrong with advanced AI systems. Both governments say they want it. Neither seems close to making it work.

The talks are happening against the backdrop of Trump’s second term and his focus on “winning the AI race” against China. According to experts who spoke to Ars Technica, that framing is already causing problems. China appears less willing to share safety intelligence when US officials treat AI development as a zero-sum competition.

The hotline proposal emerged from national security circles, not technical ones. That’s part of the problem. The people designing this communication channel aren’t the ones who would actually use it in an emergency. Technical experts from both countries, the people who could identify and explain a safety incident in real time, have largely been excluded from the planning.

What the hotline is supposed to do

The concept is straightforward. If researchers in either country discover a serious AI safety issue, something that could cause real harm or spiral out of control, there should be a fast, reliable way to alert the other side. This isn’t about espionage or intellectual property. It’s about preventing accidents that could affect both countries.

The model is nuclear weapons communication channels. During the Cold War, the US and Soviet Union maintained direct lines specifically to avoid accidental escalation. When sensors malfunctioned or military exercises were misinterpreted, those channels helped prevent catastrophic mistakes.

AI safety experts have been arguing for something similar. Advanced AI systems are being developed simultaneously in both countries. If one side discovers a serious alignment problem, a deceptive capability in a deployed model, or an unexpected failure mode, the other side needs to know. The systems they’re building likely have similar vulnerabilities.

Why it’s not happening

Trump administration officials keep framing AI development as a competition the US needs to win. That language shows up in policy documents, press releases, and diplomatic communications. It’s politically useful domestically. It’s toxic for building trust with China on safety issues.

China has been largely silent on the hotline proposal. That silence is telling. The Chinese government has its own concerns about AI safety and has implemented domestic regulations. But sharing information with a country that publicly treats you as an adversary in an “AI race” doesn’t make strategic sense.

The structural problem runs deeper. The officials negotiating this channel work on national security, not AI development. They understand crisis communication protocols. They don’t necessarily understand what information would actually matter in an AI safety incident, or how quickly that information would need to move.

There’s precedent for this failing. Nuclear safety communication worked because physicists and military officials worked together to design the protocols. The technical experts knew what signals mattered. The security officials knew how to create secure channels. Neither group could do it alone.

What happens next

The hotline isn’t dead, but it’s not close to operational. Both governments are still talking. The question is whether those talks can move beyond symbolic gestures toward something technically useful.

That requires including AI researchers and safety experts in the design process. It requires dropping the “race” framing, at least in diplomatic contexts. And it requires both sides accepting that sharing certain safety information doesn’t mean giving up competitive advantage.

The alternative is both countries developing increasingly capable AI systems without any coordination on safety issues. If something goes wrong, and we’re starting to see early warning signs that it might, there won’t be a fast way to share information about what happened or how to respond.

The people negotiating this hotline don’t need to like each other. They don’t need to trust each other’s broader intentions. They just need to agree that some AI safety incidents could affect both countries, and that having a way to communicate about those incidents is better than not having one.

Right now, we don’t have one. And based on how the talks are going, we won’t for a while.