As China mulls how to make open-weight AI less dangerous, report proposes 6-stage process

As Chinese developers and safety consultants jointly release a report focused on managing risks associated with open-weight AI models, the nation grapples with balancing innovation and safety. Open-weight AI models, which allow users to modify their underlying parameters, present a unique challenge – how to ensure safety once these models are released, given that developers relinquish control over user activity post-release.

To tackle this issue, the report proposes a six-stage process: risk identification, threshold setting, analysis, evaluation, mitigation and governance. Among the key safeguards is rigorous training-data curation, with developers urged to implement safety pre-training by filtering harmful material from training data sets before models are published.

The report's recommendations align with growing global concerns over AI safety and China's push for transparency and governance in open-weight AI development. Notably, Xiaomi streamed the reinforcement learning process for its frontier MiMo-V2.6 models, showcasing compute costs, failure logs and training-data ratios in real-time, contrasting sharply with the secretive practices of US rivals.

Chinese technology firms and regulators are also working to increase openness and oversight in open-weight AI development, with Beijing introducing a national AI Safety Governance Framework in September. As open-weight models increasingly become a focal point in the AI race between the US and China, Chinese systems have started surpassing American counterparts in key benchmarks.