Alibaba, DeepSeek defend China’s ‘misunderstood’ AI rules

Researchers from DeepSeek and Alibaba have highlighted both strengths and challenges in China’s approach to AI regulation in a recent paper published in the US journal Science.

The paper, co-authored by legal scholars and industry representatives, describes China’s regulatory framework as evolving alongside the global rise of its open-source AI systems.

Key aspects of the framework include pre-deployment filing for AI models, content safety self-assessment, exemptive rules for open-source AI and research, and phased implementation of new regulations.

The authors note that China has yet to introduce a national AI law, although proposals outlining liability for AI misuse have been circulated.

Areas needing improvement include the complexity of existing rules, and limited feedback mechanisms when filings are rejected.

The paper also warns that exemptions for open-source AI could pose risks if not carefully managed, and calls for greater transparency and independent verification mechanisms by leading Chinese AI firms.

The authors point out that China’s regulatory approach differs from that of Europe and the US, defining openness as a safety measure rather than a risk.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

China’s pre-deployment filing regime shows modest scale despite rapid registry growth

The Cyberspace Administration of China (CAC) listed 302 generative AI services by December 2024, including 238 added that year 1. The registry reached 3,739 tools from about 2,353 firms by April 2025 and grows by 250 to 300 entries each month 2.
Filing remains opaque with scant feedback on rejections, and CAC posts approvals mainly through periodic public lists 1.
Open-source exemptions exist on paper 3. The rule covers services with “public opinion attributes or social mobilization capacity”. The breadth of that trigger remains unclear, which clouds whether exemptions ease burdens for research teams or small developers.

Compliance tools and evaluation services see near-term demand from China’s detailed safety rules

China’s National Information Security Standardization Technical Committee (TC260) lists over 30 safety risks and oversight steps in its technical texts 4. Items include bias checks, supply chain security reviews, and controls on politically sensitive outputs 4. These needs open space for third-party labs and automated platforms that help with filings and audits.
Rules stress “continuous monitoring” and “regular compliance checks” 5. That sustains demand for governance tools after registration. Vendors of model evaluation software, red-teaming services (adversarial stress-testing of models), or content moderation APIs can target the current registry 2 plus new entrants each quarter.
TC260 has moved standards on labels for AI-generated content, with some already in effect 6. That creates near-term openings for firms that ship tech aligned with these label rules.