SemiAnalysis: 3.6% of Chinese model releases publish safety results
Confirmed
The takeaway
On Oct 8, 2026 SemiAnalysis (Dylan Patel and two co-authors) published a reply to Dario Amodei’s “pace the frontier” essay.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 9 October 2026
Your AI and this story
- GPT-6 Astra161 days after its cutoff
- Claude Opus 5.5100 days after its cutoff
- Gemini 3.8 Flash191 days after its cutoff
- Grok 4.7130 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 100 days before it.
Key facts
- Dataset: 857 releases (2021 – Sept 2026) from ByteDance, Alibaba, Tencent, Baidu, DeepSeek, Moonshot, Zhipu, MiniMax and StepFun; 31 (3.6%) with any published safety result, 9 (1.1%) at or before launch, 813 (94.9%) with no safety disclosure; no policy milestone changed the rate
- Per lab: Alibaba 238 releases, 7 with safety results (3 at launch); Tencent 133 (1); ByteDance 120 (2); Baidu 49 (1). Zhipu (Z.ai) is the only developer with consistent yearly safety documentation
- Policy: China’s AI Safety Governance Framework 3.0 (Sept 2026) names loss-of-control risks such as deception and capability hiding but calls ‘promoting AI innovation and development’ the first priority; the comprehensive AI Law was shelved in 2025; there are no compute thresholds that trigger frontier-risk duties (unlike the EU’s 10^25 FLOP or California SB 53’s 10^26)
- Expert survey: 102 Chinese documents (2023 – Sept 2026); 86% of technical/scientific texts mention frontier or loss-of-control risk vs 22% of legal texts; safety bodies such as Shanghai AI Lab, CAICT, CnAISDA and TC260 have no power to compel testing
- Leaders: SemiAnalysis finds no public safety statement from DeepSeek’s Liang Wenfeng in four years and none on safety in a 6,000-word superintelligence keynote by Alibaba’s Eddie Wu; Moonshot’s Yang Zhilin acknowledged risks but chose acceleration
- It also points out that Anthropic shipped Claude Opus 5.5 ten days after Amodei’s Sept 12 pacing essay, framed partly as keeping an edge over China
What happened
SemiAnalysis published “Beijing will not pace the frontier” on October 8, 2026 (SemiAnalysis1), answering Dario Amodei’s September call for frontier labs to slow down together. The authors coded every model release they could find from nine leading Chinese developers since 2021 and checked whether safety results were ever published. Out of 857 releases, 31 had any, and 9 had them at launch. Zhipu was the only company with regular safety reporting.
The report also reviews Chinese policy and expert writing. Official frameworks now use loss-of-control language, but binding rules apply to content and public-facing services, not to frontier capabilities, and officials consistently put development first. Chinese technical experts are far more worried about frontier risk than legal and policy writers, but the bodies that study safety cannot compel testing.
Why it matters
“Pacing the frontier” depends on China slowing too. This is the most detailed public dataset so far on how rarely Chinese labs publish safety evaluations, and it gives critics of unilateral slowdowns a concrete argument. It is an analysis by a commercial research firm, not an official audit.
Sources
1 source from 1 site. Numbers match the chips in the text.
1 source: 1 reaction
Reactions
- SemiAnalysis: Beijing will not pace the frontiernewsletter.semianalysis.com, discussion
Changes
- Filed