Chinese AI Labs Disclosed Safety Evals for Just 3.6% of 857 Model Releases

A SemiAnalysis review of 857 model releases from nine leading Chinese AI companies found only 31 (3.6%) carried a safety-evaluation result that could be matched to a specific model — and just nine published one at launch. The pattern points to a speed-first regulatory regime, Reuters reports.

Key Takeaways

A SemiAnalysis review of 857 model releases from nine leading Chinese AI companies found only 31 (3.6%) carried a safety-evaluation result that could be matched to a specific model — and just nine published one at launch.
The pattern points to a speed-first regulatory regime, Reuters reports.

AI synthesis4 sources1

Chinese AI developers published safety-evaluation results for just 3.6% of the models they released over the past five years, according to a new review that casts doubt on how much scrutiny the world's second-largest AI ecosystem applies to its own frontier models.

The California-based research firm SemiAnalysis examined 857 model releases from nine leading Chinese AI companies between 2021 and September 15, 2026, Reuters reported. Only 31 releases — 3.6% — came with a safety-evaluation result that could be matched to a specific model, and just nine of those (1.1%) had such results published at launch. That leaves 813 releases with no model-specific safety disclosure at all.

The nine companies include four hyperscalers — ByteDance, Alibaba, Tencent and Baidu — and five startups. SemiAnalysis described the pattern as an outcome of a 'speed-first' regulatory architecture: in its reading, Beijing's rules do not force labs to slow down for disclosure, and the incentives point toward shipping first.

The finding lands in a busy week for AI safety politics. Microsoft chief executive Satya Nadella argued that every advanced model should be assumed 'compromised' and be subject to containment and audit, with an emergency brake any authorised person can pull. The White House, meanwhile, moved to require AI firms to report incidents immediately.

None of this proves the untested models are unsafe, and Chinese developers do publish some safety research; several also see disclosure requirements as tools of Western competitive pressure. But the gap between the volume of frontier releases and the near-absence of model-specific safety disclosure is exactly the kind of asymmetry regulators in Washington and Brussels say they want to close — and it is harder to close when the releases outpace the paperwork by two orders of magnitude.