The 'But China' Argument Against US AI Safety Is Hiding in Two Companies

The "but China" talking point ends roughly half of US AI policy arguments. The claim is that any safety rule is a gift to Beijing because China doesn't slow down. State legislators hear it. The 2026 federal preemption fight hears it. The argument treats China as a fixed policy variable and US safety rules as pure downside. It is also testable against public evidence, and the evidence is unflattering to the talking point.

The clearest test comes from Concordia AI's airiskmonitor.net, a China-based independent benchmark that plots capability scores against safety scores for 47 frontier models. The scatterplot covers five risk axes: cyber, bio, chem, manipulation, and loss-of-control. Concordia is a Chinese organization publishing the comparison, which is part of the point. The two highest-safety American entries in the plot are OpenAI and Anthropic. Take them out, and the remaining US field clusters much closer to the Chinese cluster than the "but China" framing suggests.

Adam Gleave of FAR.AI independently lands on the same ordering through jailbreak-robustness work. OpenAI and Anthropic are the hardest to break. Google's Gemini and xAI's Grok are easier. Chinese models are easier still, but the gap between those and the rest of the American field is much smaller than the gap between the duopoly and the rest of the American field. Two independent measurement efforts, different methodologies, same shape. The duopoly is doing the work.

That synthesis comes from Nathan Labenz, host of the Cognitive Revolution podcast, who spent two weeks in China in July 2026 attending the World AI Conference (WAIC) in Shanghai and an AI safety hub launch in Beijing, recording under Chatham House rules, a convention in which remarks are reported without naming the speaker. Chatham House status means private conversations are not quotable, so the load-bearing evidence is the public artifacts: Concordia AI's scatterplot, FAR.AI's jailbreak-robustness work, and a "45-degree line" framing Labenz encountered in Beijing, where Chinese researchers argue capability and safety should rise together rather than trade off.

The Chinese side is not winning. Labenz's stated concession, on the record, is that Chinese models and services still do not have safeguards as strong as the leading American ones. The "45-degree line" framing treats that as a target rather than an achievement, and Chinese AI safety work sits inside universities rather than independent labs, with rapid growth in research on self-replication, deception, evaluation faking, and interpretability, the techniques for understanding what a model is doing internally. The convergence story is real. The lead story is also real. The duopoly claim is the missing middle.

The policy lever follows from the measurement. If US safety rules are written to avoid handing China the lead, they are regulating against a phantom competitor. The actual risk that state-level safety laws and the 2026 federal preemption fight are trying to manage sits inside two companies. The rest of the field, American and Chinese, is closer than the talking point allows, which means the policy unit is the duopoly rather than the country.

The next test is whether the federal preemption debate names that. If preemption moves forward on a "but China" rationale, the rule it preempts is one that would bind the duopoly more than anyone else, which is also the rule the duopoly's own safety work already exceeds. The talking point and the evidence are pointed in different directions.