Fact-Checking China's AI Model Restrictions: What the Supreme People's Court Actually Says About Open-Source Governance

Beyond the Headlines

In early July 2026, Reuters published reports claiming that China is curbing access to its top AI models, citing unnamed sources and suggesting restrictive governance measures. These reports were repeated across LinkedIn and other channels, so I wanted to check it out.

If true, it would be a significant concern in the global AI community, particularly among organisations relying on open-source alternatives like Alibaba's Qwen class models.

The best way to fact-check was going directly to the source—examining the actual documents from China's Supreme People's Court and the official roundtable discussions on AI governance. What we found tells a more nuanced story than the Reuters reporting suggests—one focused on differentiated regulation, ecosystem coordination, and responsible innovation rather than blanket restrictions.

Why This Matters

Many organisations are starting to adopt running open-source models like Alibaba's Qwen as cost-effective alternatives to buying AI services from USA-based frontier model providers like OpenAI or Anthropic. For startups, research institutions, and increasingly enterprises, open-source models democratise access to cutting-edge AI capabilities.

These open-source models are sometimes used 'as is' as a direct and cheaper alternative to the frontier models, but in other cases the open-source models are a platform to domain-specific variants of models with custom weights and fine-tuning to match the organisations' own domain-specific use-cases (which is a pattern we advocate for by the way—a topic for another paper!).

Any regulatory action that restricts this access has global implications and reinforces the business case for countries and major organisations to invest in their own Sovereign capabilities. As we saw in June when the USA government imposed trade restrictions on Anthropic—impacting availability to the latest Fable/Mythos models—geo-political factors and supplier concentration risks are a live and very real risk that anyone depending on AI now needs to be taking extremely seriously.

Reuters Claim vs. Reality: What the Documents Actually Say

The Reuters Narrative

Reuters reported that Chinese authorities are "looking at curbing overseas access" to top AI models and that this represents a shift toward more restrictive governance. The implication was that China is moving toward gatekeeping its AI capabilities.

What the Chinese Supreme People's Court Documents Reveal

The actual Supreme People's Court roundtable discussion on AI governance presents a far more sophisticated framework. Rather than blanket restrictions, the documents outline a differentiated approach that distinguishes between various types of AI systems, use cases, and risk profiles. The emphasis is on "layering", "coordination", and "responsiveness"—not prohibition.

The key distinction: the Chinese approach is not about restricting access to open-source models, but rather about ensuring that AI systems—whether open or closed—operate within a framework of accountability, transparency, and risk management.

The Actual Chinese Governance Approach: A Differentiated Framework

Based on the Supreme People's Court roundtable discussion, China's AI governance strategy rests on four core principles:

  • Differentiation: Regulations distinguish between frontier models, open-source models, fine-tuned systems, and application-layer AI. Not all AI systems face the same requirements.
  • Layering: Governance operates at multiple levels—national, sectoral, and organisational. Different layers have different responsibilities and authorities.
  • Coordination: Government agencies, industry bodies, and technology companies work together to develop standards and best practices rather than through top-down mandates alone.
  • Responsiveness: The regulatory framework is designed to evolve as technology and risks change, avoiding rigid rules that quickly become obsolete.

This framework is notably different from either a "free market" approach or a "blanket ban" approach. It acknowledges that different AI systems pose different risks and require different governance mechanisms.

Competition Risks: Ecosystem Lock-In and the Open-to-Closed Conversion Pattern

While the Chinese governance framework is more nuanced than Reuters suggested, there are legitimate concerns about competition dynamics in the AI ecosystem. Three risks deserve your attention:

  • Ecosystem Lock-In: As organisations invest in fine-tuning and customising open-source models, they may become dependent on specific model architectures, training data characteristics, and vendor ecosystems. This creates switching costs that can limit competition.
  • Gatekeeping: Providers of open-source models could theoretically use their position to favour certain downstream applications or restrict access to specific use cases, creating a form of gatekeeping that appears less obvious than closed-source restrictions.
  • Open-Source-First-Then-Closed-Source Conversion: A concerning pattern has emerged where companies release models as open-source to build adoption and ecosystem lock-in, then gradually transition to closed-source or restricted-access models as they gain market dominance. Or it's commonplace for the latest models to be only available initially via the commercial closed-source route.
  • The Irreversibility Problem: Once an AI model is released as open-source, it cannot be "un-released." The model weights, architecture, and training methodology become part of the global commons. This creates a fundamental asymmetry: governments can restrict future releases, but they cannot control what has already been distributed.

These risks are not unique to China—they are inherent to how technology ecosystems evolve. However, they highlight why governance frameworks that maintain transparency, prevent anti-competitive behaviour, and preserve genuine open-source alternatives are important globally.

QuivaWorks Perspective: Open-Source Models in a Regulated Landscape

QuivaWorks AI operations platform is model/provider agnostic—for any given AI Assistant you can plug-and-play your chosen model as best suits the use-case in mind.

Out of the box if you sign up online, we make Anthropic's suite of models available (we've found them to be best commercially available for agentic use), but we do also have deep expertise in building and running our own models both from scratch and using open-source models.

For enterprise and government clients we offer access to custom domain-specific models. It's a topic for another post but we have some sophisticated multi-model orchestration and validation infrastructure which means if you want to use AI for domain-specific purposes then we can help you get better performance, at significant cost reduction compared to frontier models, and it's fully Sovereign too (your models and your infrastructure).

Anyway, based on our experience, some considerations…

Building Foundation Models from Scratch

We are one of the estimated <0.1% of teams that have built a foundation model from scratch. This approach provides maximum control over training data, model architecture, and performance characteristics. But it requires enormous computational resources and expertise. Building our own foundation model gave us a top-performing model for our intended use-cases with full governance controls for the tightly regulated environments our clients work within.

Leveraging Open-Source Models

We also have used open-source models as foundations for domain-specific applications. This approach offers several compelling advantages:

  • Accelerated Development: Starting with a pre-trained model reduces time-to-market significantly. Rather than training from scratch, we focus on fine-tuning and optimisation.
  • Leverage Mass Datasets: Depending on the model you choose, some open-source models can be trained on enormous, diverse datasets which would be impractical or at least economically unviable for us to build ourselves. They can give you access to a strong foundation that captures broad patterns in language and reasoning.
  • Cost Efficiency: You avoid the massive computational costs of training frontier models from scratch, making advanced AI accessible if you've got a more constrained budget.

But using open-source models has disadvantages, or things you need to be aware of too:

  • Inbuilt Bias: Someone else's model weights may carry significant biases from their training data. Based on our experience, these biases are not merely statistical artifacts—they can reflect real patterns in the data that can perpetuate or amplify societal inequities. We've previously had to spent considerable time stripping out bias from open-source models before we could use them confidently.

Conclusions

In conclusion, it seems that the reports of China's planned prohibitions were over enthusiastic.

Nevertheless, there is more governance pressure and oversight coming and this reinforces the criticality of making detailed risk assessments as part of your model governance activities.

Open-source models are likely here to stay as part of the ecosystem but it's important that you take seriously the vendor and geopolitical risks, that you understand what data the models you're relying on were trained upon and any biases that the model weights may contain.

If you're interested in running your own models, and how you can optimise performance for your specific use-cases at significant cost reduction—please reach out, we'll be happy to offer a briefing and demo of our proprietary multi-model orchestration and validation tools.

Sources and References