Updated
Updated · The Guardian · Aug 7
Zhejiang Lab’s GeoGPT Fails UN Openness Test, Releasing 0 Training Data
Updated
Updated · The Guardian · Aug 7

Zhejiang Lab’s GeoGPT Fails UN Openness Test, Releasing 0 Training Data

1 articles · Updated · The Guardian · Aug 7

Summary

  • GeoGPT, a geoscience AI system promoted by Zhejiang Lab as open, would not qualify as open under a UN-endorsed model openness framework, according to the latest scrutiny of China’s AI transparency claims.
  • 0 training data and no application source code are released, while the model relies mainly on Alibaba’s Qwen weights and a governance committee that answers to Zhejiang Lab itself.
  • ESA’s Earth Virtual Expert publishes training and evaluation tooling, and Ai2’s OLMo goes further by releasing the full materials needed for reproducibility, highlighting the gap with GeoGPT.
  • That distinction matters because a free hosted service is not necessarily open: GeoGPT users route data through China’s legal jurisdiction, creating dependency and sovereignty risks.
  • The critique argues the issue is broader than China alone and says any UK-China AI cooperation should be tied to shared UN-style openness standards rather than diplomatic claims.

Insights

Are heavily promoted open AI models secretly compromising your data sovereignty by hiding critical jurisdictional loopholes?
If healthcare AI relies on opaque systems, who truly holds the ultimate power over life-or-death medical decisions?