Report Finds Limited Public AI Safety Testing Disclosures by Chinese Developers

Computer-generated English audio

China News | October 10, 2026. A research review has raised questions about how much information Chinese artificial intelligence developers publicly release about model safety testing. The findings concern transparency of published evaluations, not definitive proof that undisclosed tests never took place.

What the research examined

SemiAnalysis reviewed 857 model releases from nine Chinese AI developers between 2021 and September 15, 2026. The review covered companies including Alibaba, ByteDance, Tencent, Baidu and DeepSeek.

The headline finding

Researchers identified 31 releases with public safety-evaluation results tied to a particular model. That represents 3.6% of the releases reviewed. Only nine had those results publicly available by launch, according to the report.

What the percentages do and do not mean

The analysis measures publicly available, model-specific disclosure under the researchers’ criteria. A release without a published evaluation may still have undergone internal testing. Conversely, publication of test results does not by itself guarantee that a model is safe in every setting.

Why disclosure matters

Public documentation can help independent researchers, enterprise customers and policymakers understand a system’s limitations and compare risk-management approaches. Differences in evaluation methods and reporting standards can make comparisons difficult.

The broader AI policy context

China is simultaneously encouraging domestic AI development and discussing stronger risk monitoring. Future disclosures and independent verification will help show whether safety reporting becomes more consistent.