SemiAnalysis: Just 1.1% of Chinese AI Models Share Safety Data at Launch, Review Finds
A review of 857 model releases from nine Chinese labs reveals minimal safety transparency, with only 1.1% sharing results at launch.
Key points
- SemiAnalysis reviewed 857 AI releases from nine Chinese labs between 2021 and September 2026.
- Just 3.6% of these releases included developer-disclosed safety results.
- Only 1.1% of releases provided safety data at the time of initial launch.
SemiAnalysis published a study showing that Chinese AI laboratories rarely disclose safety evaluation results for their model releases. The analysis covers 857 releases from nine labs over a six-year period, revealing significant gaps in transparency.
What happened
SemiAnalysis conducted an analysis of 857 artificial intelligence releases from nine Chinese laboratories between 2021 and September 2026. The study found that only 3.6% of these releases included safety results disclosed by the developer. This low figure indicates a widespread lack of public safety reporting in the sector.
The data shows that transparency was even lower at the point of release. According to SemiAnalysis, only 1.1% of the reviewed models provided safety results at launch. Most laboratories did not share safety evaluations until after the model was already available to users or developers.
This reporting comes shortly after a significant industry statement on safety pacing. On 12 September 2026, Dario Amodei published an essay arguing that frontier labs must deliberately slow the pace of improving model capabilities. Amodei’s argument highlights growing concerns about the rapid advancement of AI systems without corresponding safety measures.
Why it matters
The lack of safety disclosures creates significant risk for organizations deploying these models. IT managers and CISOs cannot accurately assess the potential dangers of a model if the developer does not provide safety data. This opacity makes it difficult to implement effective risk mitigation strategies before deployment.
The timing of these findings aligns with broader industry debates on AI safety. Dario Amodei’s September 2026 essay suggests that the speed of capability improvements outpaces safety development. The SemiAnalysis data supports the concern that many labs are prioritizing release volume over safety verification.
For global technology firms, this trend complicates compliance and security planning. If a model lacks published safety results, organizations must assume higher uncertainty regarding its behavior. This uncertainty can lead to unexpected vulnerabilities or misuse in production environments.
The disparity between launch-time transparency and post-launch disclosures also affects trust. Users and partners may struggle to verify the safety claims of a model if data is released retroactively. This delay can hinder the ability to conduct thorough independent audits before integration.
As the AI landscape evolves, the pressure on labs to demonstrate safety will likely increase. The current low disclosure rates suggest that voluntary transparency is insufficient. Regulatory bodies and industry groups may need to enforce stricter reporting standards to ensure consistent safety practices.
What to watch
- Monitor for new regulatory requirements on AI safety disclosures.
- Track industry responses to Dario Amodei’s essay on slowing model development.
- Watch for updates from the nine Chinese labs included in the SemiAnalysis study.
- Observe trends in post-launch safety reporting versus launch-time transparency.
What to do and how to stay safe: SemiAnalysis
- Review current AI vendor contracts to ensure they include clauses for safety reporting.
- Develop internal protocols for assessing models that lack public safety data.
- Stay informed about industry standards for AI safety and transparency.
- Consult with legal teams regarding liability when using models with undisclosed safety results.
Step-by-step guide: Software Updates Best Practices: Secure Patching Without Downtime
General security guidance from the Patch Gazette newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How many AI releases did SemiAnalysis review?
SemiAnalysis reviewed 857 releases from nine Chinese AI labs between 2021 and September 2026.
What percentage of releases disclosed safety results at launch?
Only 1.1% of the releases included safety results from the developer at the time of launch.
Who argued for slowing AI model development in September 2026?
Dario Amodei published an essay on 12 September 2026 arguing that frontier labs must deliberately slow the pace of improving model capabilities.
