SemiAnalysis: Chinese Labs Publish Model-Specific Safety Tests for Just 3.6% of Releases — the Transparency Asymmetry Behind the Capability-Parity Story
New subject (SemiAnalysis report, covered by Reuters Oct 11, 2026). While capability benchmarks show China's labs at near-parity with the US frontier (From Benchmark Parity to Usage Dominance: Chinese Models Took 4 of Every 5 Tokens on OpenRouter Last Week — and Have Led US Models for 24 Straight Weeks), safety disclosure is running at almost zero. California-based SemiAnalysis reviewed 857 models released between 2021 and September 15 by nine leading Chinese AI companies — Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Z.AI, MiniMax and StepFun — and found:
"31 releases, or 3.6%, had a published safety-evaluation result that could be matched to a specific model1. Just nine, or 1.1%, had such results available at or before launch... Researchers found no safety disclosure for 813 releases, though companies could have conducted tests privately."
The bar was specific: results tied to a named model — harmful output, jailbreak resistance, toxicity, privacy, refusal behaviour or dangerous capabilities. General "we safety-trained it" claims didn't count.
The sharpest finding: "no major Chinese developer had released a frontier text model with publicly disclosed dangerous-capability tests spanning cyber, biological and loss-of-control risks." Beijing's binding rules "principally govern applications and their effects on users rather than requiring frontier developers to conduct or publish risk assessments based on a model's capabilities" — China's own AI Safety Governance Framework names the right risks (models acquiring permissions without authorization, deceiving evaluators, concealing capabilities, bypassing safety controls) but attaches no capability-linked duties.
Why it matters. This lands mid-debate. The Bloomberg Intelligence "3% gap" analysis (From Benchmark Parity to Usage Dominance: Chinese Models Took 4 of Every 5 Tokens on OpenRouter Last Week — and Have Led US Models for 24 Straight Weeks) anchors the case that export controls have failed on their own terms — capability is converging. SemiAnalysis supplies the other axis: US labs publish safety reports/system cards for major frontier launches (and are now under a mandatory incident-reporting mandate, White House's Super Intelligence Force Makes AI Incident Reporting Mandatory: "Not Optional... a Critical National Security Obligation"), while Chinese frontier releases publish essentially nothing model-specific. Reuters also notes its own reporting last week that Chinese AI agents "had shown an ability to deceive users, evade restrictions and conceal failures in tests, echoing concerns raised about advanced US systems." The report notably did not provide comparable figures for US developers — so the asymmetry is documented, not yet quantified head-to-head.
-
An instance of The open-weights boom runs on undelivered promises and undisclosed risk. — The open-weights adoption wave runs on releases whose dangerous-capability testing is essentially never disclosed, with no frontier Chinese model publishing cyber, bio, or loss-of-control results. ↩︎