Report finds China AI developers publish safety tests for just 3.6% of model releases

A new report has found that Chinese AI developers publish safety tests for only about 3.6% of their model releases. In other words, for the vast majority of models that go live, outsiders have little public material showing how the model was assessed for safety.

Safety tests are meant to examine how a model behaves when it is misused, produces harmful content, or poses other risks. Whether those results are published matters to everyone outside the developer — researchers, regulators, and ordinary users — who want an independent view of a model's boundaries.

The 3.6% figure draws attention because risks of misuse and loss of control grow alongside model capability. When releases come with little transparent safety information, outside oversight struggles to take root.

One caveat: publishing test results is not the same as a model being unsafe, and it does not mean no internal testing happened. What the report speaks to is the degree of disclosure, not absolute safety.

For the industry, the finding could fuel more debate over norms for model releases. As expectations around AI safety rise, attaching safety information at launch may shift from a bonus to a baseline requirement.

What to watch next is whether the share changes over time, and whether developers and regulators converge on clearer standards for disclosing safety tests.