US Techies Embrace Chinese AI as Safety Checks Lag

What's happening

A growing number of American technologists building products and workflows on top of AI are quietly dropping national origin as a filter for which models they use. Mozilla CTO Raffi Krikorian says he relies on Kimi K3, a Chinese-built model, for much of his daily work, and describes the open frontier of AI development as increasingly Chinese in character 1. That attitude — pick whatever model performs best and is cheapest to run, regardless of where it was built — is emerging just as the systems meant to check whether these models are safe are struggling to keep pace 2.

The backdrop is a frontier AI field that is releasing capable new models at a pace that outstrips the infrastructure built to evaluate them. Researchers tasked with testing frontier systems for dangerous capabilities say rising compute costs and shrinking release windows are squeezing their ability to do rigorous evaluation, even as the models themselves become harder to probe and measure 2. In just a two-week span, five separate frontier-class models were released — Claude Fable 5, Grok 4.5, GPT-5.6 Sol, Muse Spark 1.1, and Kimi K3 — a cadence one industry account frames as evidence of an accelerating competitive race among AI labs with direct consequences for cybersecurity defense and misuse risk 5.

Policymakers and industry figures are proposing different fixes for the same underlying worry: that frontier models are shipping faster than anyone can verify they're safe. In Washington, a proposed bill would give the Department of Homeland Security authority to order AI developers to slow down or shut off frontier models following a catastrophic incident, an approach its own proponents acknowledge could create serious headaches for enterprises that depend on those systems running continuously 3. Elon Musk, in an interview with The Economist, took a different tack, calling on major AI companies to peer-review each other's most advanced models before public release, arguing that competitive pressure has outrun any meaningful external check on what's being shipped 4.

Where the reporting agrees

Across this coverage, there's a consistent thread: frontier AI development is moving faster than the mechanisms meant to govern it. The evaluators charged with catching dangerous capabilities before release say they're falling behind the pace of new model launches 2, the tally of five major releases in two weeks is offered as concrete evidence of that acceleration 5, and both the proposed kill-switch legislation 3 and Musk's peer-review proposal 4 are explicitly framed as responses to the same worry — that nobody is meaningfully checking these systems before they reach the public. Even the Fortune reporting on American engineers adopting Chinese models fits this pattern indirectly: if practitioners are choosing tools based on capability and cost rather than provenance or vetted safety pedigree, that reinforces the picture of a market moving faster than formal oversight can track 1.

Where it doesn't

The sources diverge sharply on what kind of intervention is appropriate, and on how much weight to put on any single fix. The House bill described by TechRepublic is a regulatory, government-enforced mechanism — DHS-ordered shutdowns after the fact — that critics worry could disrupt businesses relying on these models in production 3. Musk's proposal, by contrast, is voluntary and industry-driven, asking labs to check each other's work before release rather than waiting for a regulator to intervene after a catastrophe 4. These aren't just different tactics; they reflect different theories of where the failure lies — one assumes government must backstop industry, the other assumes industry can and should police itself.

There's also a gap in how directly sources connect Chinese model adoption to the safety conversation. Fortune's reporting treats the rise of Chinese-built open models like Kimi K3 as a market and geopolitical story about where frontier development is happening 1, while the cybersecurity-focused rundown of recent releases lists Kimi K3 alongside Western models purely as part of a rapid-release trend with security implications, without addressing its Chinese origin as a distinct factor 5. Neither ties the adoption trend explicitly to the evaluators' capacity problem described elsewhere 2, leaving that connection as inference rather than something the reporting itself asserts.

The likely reading

Taken together, the evidence points toward a widening gap between deployment speed and evaluation capacity, with policy responses still scattered and unproven. The kill-switch bill and the peer-review proposal are best read not as competing solutions already in motion, but as early, untested reactions to a problem that the evaluators themselves say they can't yet solve with existing tools 234. Meanwhile, practitioners voting with their workflows — adopting whichever model works best, Chinese or not — suggests the market is settling questions of adoption long before regulators or labs settle questions of safety 15.