China weighs ban on open-weight AI exports amid distillation fears

Chinese authorities have held talks with Alibaba, ByteDance, and Z.ai about potentially restricting foreign access to their most advanced AI models—including those not yet released, according to Reuters. This would mark a dramatic U-turn from China's current open-weight approach, where model weights are published online for anyone to download and run. That openness has been China's key advantage in global AI adoption, allowing companies like Alibaba and Z.ai to compete despite models trailing America's best by roughly seven months on key benchmarks. Options being discussed include barring public release or limiting models to domestic use only, though no official decision has been made.

At the heart of the debate is "distillation"—using outputs from smarter AI models (like Anthropic's Claude or OpenAI's GPT-5.6) to improve weaker ones. Anthropic has accused Chinese labs of systematically distilling Claude's capabilities, citing 16 million exchanges from 24,000 fraudulent accounts. The company added hidden code to Claude Code to detect Chinese users, which led Alibaba to ban the tool internally. Anthropic frames distillation as a national security threat, warning that distilled capabilities could enable military and surveillance systems. Meanwhile, Washington has already restricted access to Anthropic's Mythos and OpenAI's GPT-5.6 over similar security concerns, suggesting a coordinated front may emerge.