AI · SEMICONDUCTOR · FACT CHECK
SKD WIREMostly True
Only a small fraction of AI models developed in China disclose their safety evaluations, and disclosures made before release are even rarer, according to findings reviewed against primary sources — with the overall claim assessed as largely accurate, subject to minor caveats.
The claim centers on two figures: that 3.6% of Chinese AI models have made safety evaluation results public, and that only 1.1% disclosed such evaluations before their release. Both figures were found to align with the underlying material, including documentation available through a public Yahoo Tech mirror of the report.
The figures point to a broad gap between the growing number of AI models released in China and the transparency practices surrounding their safety testing. Pre-release disclosure — publication of evaluation results before a model becomes publicly available — is markedly less common than disclosure at any point.
The assessment carries some uncertainty. The determination was made by substituting content via a Yahoo Tech public mirror, meaning the review relied on an alternate access route to the source material. The exact population of Chinese AI models covered and the methodology behind the percentages were not fully detailed in the materials reviewed, so the figures should be understood within that context.
No additional primary documentation was available to clarify whether the percentages apply to a specific reporting period or to all Chinese models to date.
The core figures — 3.6% of Chinese AI models publishing safety evaluations and only 1.1% doing so before release — are supported by the sources reviewed, with minor reservations about the access route and scope. On balance, the claim is Mostly True.
