Chinese labs are wild cards in AI governance

Alibaba Holding

30/09/2026 13:00

The author is a Reuters Breakingviews columnist.  The opinions expressed are her own.
By Robyn Mak
HONG KONG, Sept 30 (Reuters Breakingviews) -  Artificial intelligence bots from OpenAI and Anthropic going off-script have exposed alarming shortcomings in how the two manage and disclose catastrophic risks. But Chinese firms from Alibaba 9988.HK to Z.AI 2513.HK are less transparent in AI safety than US peers. They'll have to up their game to retain access to foreign markets from Brussels to Seoul.
     As the world frets over the havoc rogue agents could unleash, the $38 billion Z.AI on Monday quietly released its maiden AI safety framework - the first from a major Chinese model developer. Whereas OpenAI and Anthropic routinely publish model assessments, such voluntary disclosures are rare from companies in the People's Republic, where only half of the ten leading model developers reported safety evaluation results alongside releases in the past year, according to a July report from consultancy Concordia AI.
Z.AI's latest framework is welcome but more can be done. In August, it delayed releasing its flagship GLM-5.3 model by two weeks to conduct safety tests and strengthen safeguards. But it didn't disclose what it fixed during the delay, or release any broader safety evaluations beyond hacking-related risks. By comparison, Anthropic has a public 212-page "system card" detailing the capabilities, limitations and risks of its restricted Mythos 5.1 and Fable 5.1 models.
One reason Chinese offerings lag in safety disclosures is that they have only started to near US-level frontier capabilities, prompting companies and authorities to begin paying greater attention to the accompanying risks. A Reuters analysis of more than 200 research documents and technical reports found at least two dozen examples of Chinese AI-powered agents learning to deceive and scheme and challenge boundaries since last year. Another factor is that AI firms in the People's Republic face a shortage of chips and capital, and running safety tests can be expensive and requires computing power.
But as their sophisticated open-weight models gain traction in overseas markets, Chinese companies will have to at least match their US rivals in transparency and governance. In Europe, for example, the EU Artificial Intelligence Act will require certain "high-risk" AI providers to establish risk and quality management systems, perform model evaluations, among other things, by next year, though that time line has already been delayed. Similarly, "high-impact AI" systems in South Korea are subject to comprehensive safety and reliability measures.
Silicon Valley labs are no saints, as underscored by Australia's lambasting of Sam Altman this month after a rogue OpenAI agent hacked into government websites. But while the US may remain ahead in the race for frontier dominance, Chinese labs have a chance to close the governance gap at speed.
Follow Robyn Mak on X.
     
CONTEXT NEWS
Chinese artificial intelligence company Z.AI and Beijing-based consultancy Concordia AI on September 28 released an open-weight model risk management report, the first comprehensive AI safety framework from a Chinese company.
Separately, Z.AI on September 21 said it had disabled some features of its flagship AI coding assistant ​after some users reported it was uploading entire local code repositories onto overseas cloud servers without their consent.
In August the company released its flagship GLM-5.3, which it says nears Anthropic's Mythos in finding software vulnerabilities, after a two-week review, becoming the first Chinese lab to explicitly delay an AI model release for safety reasons.

(Editing by Una Galani; Production by Aditya Srivastav)
((For previous columns by the author, Reuters customers can click on MAK/  robyn.mak@thomsonreuters.com))