Study finds safety results published for 3.6% of Chinese AI model releases
SemiAnalysis matched public safety-evaluation results to 31 of 857 model releases by nine Chinese developers; the study does not establish whether other models were tested privately.
Public results were rare in the sample
SemiAnalysis says it reviewed 857 identifiable model releases by nine major Chinese developers from 2021 through September 15, 2026. It found a published safety-evaluation result tied to a specific model for 31 releases, or 3.6%. Results were available at or before launch for nine releases, or 1.1% of the total. Reuters reported the findings on October 9. The researchers counted substantive results for areas such as harmful output, jailbreak resistance, privacy or dangerous capabilities; a general claim that a model had been evaluated did not qualify.
What the figures do and do not show
The study examined public materials, so the absence of a matching result does not show that a developer skipped internal testing. Its release count includes model sizes and snapshots, which can affect comparisons between companies. SemiAnalysis also says none of the nine developers had routinely disclosed the dangerous-capability tests it sought for frontier text models. Reuters notes that the report did not calculate comparable disclosure rates for US developers. The findings describe a transparency gap within this defined sample, not a direct measurement of every model’s safety or a cross-country safety ranking.