Ben Luria, co-founder and chief executive of AI safety startup Hirundo, said his team edits model weights to realign "state-aligned bias and propaganda" and other political framing to what Western companies expect from the systems they deploy. The work also hardens models against cyberattacks, increases fairness and reduces dangerous knowledge such as information on bioweapons. "We're not making the model anti-Chinese," he said. "We are bringing it up to speed on safety and to the standard of mainstream Western AI."

The concerns have some supporting evidence. CrowdStrike's analysis of DeepSeek's R1 found prompts touching politically sensitive topics were up to 50% more likely to receive code with severe security vulnerabilities, and Hirundo's own review of Alibaba's Qwen models found they would not engage on subjects such as the 1989 Tiananmen Square crackdown or Xinjiang. Boston-based Lazarus AI, which realigned Qwen models for regulated industries, found that after removing the censorship the models could produce information absent from their training materials, such as names of pro-democracy activists, suggesting the knowledge was learned but suppressed from expression.

The commercial stakes are real: top Chinese models trail leading US closed systems by just 3% on benchmark scores and account for the biggest share of downloads on Hugging Face, and companies including Airbnb, Uber and legal AI firm Harvey have built proprietary models on top of them. But doubts persist. In October, the Trump administration barred civilian government agencies from using DeepSeek products, citing potential national security risks.