China's AI openness claims don't hold up under scrutiny

Key points

• China's GeoGPT, a geoscience AI system, was showcased at the World AI Conference in July 2025 as an example of open, jointly governed AI.
• Under the "model openness framework" endorsed in a recent UN report, GeoGPT would not qualify as open.
• GeoGPT's underlying model weights are built mainly on Alibaba's Qwen, whose licences do not meet the Open Systems Interconnection standard for true openness.
• Neither GeoGPT's training data nor its application source code has been released publicly.
• Its governance committee reports to Zhejiang Lab itself, not to an independent body.

China's AI labs have genuinely given the world some impressive, freely available tools. Qwen, DeepSeek and Kimi, three large language models (the technology that powers chatbots like ChatGPT), all come from Chinese teams and have been released in ways that let developers worldwide download and run them. For researchers in the developing world working on modest computers, that matters a great deal.

But "openly released" and "truly open" are not the same thing. That distinction is at the heart of a debate playing out in the pages of The Guardian and beyond.

What is GeoGPT and why does it matter?

GeoGPT is an AI system built for geoscience research by Zhejiang Lab, a Chinese state-backed research institute. Last month, at the World AI Conference, it was held up as a model of open, jointly governed science. Governments in the developing world were invited to treat it as a shared resource.

The problem: it doesn't meet the standard definition of "open" that independent experts and a recent UN report have endorsed.

Under the model openness framework, a genuinely open AI system must release its model weights (the internal settings that make the AI work), its training data, and its source code. GeoGPT releases only the weights, and even those come with strings attached: they sit on top of Alibaba's Qwen model, which uses licences that don't comply with the Open Systems Interconnection standard, a widely recognised benchmark for open software. No training data. No source code. And the committee that supposedly governs the project answers to Zhejiang Lab itself.

That is less "open science" and more "open-ish, on our terms".

Is this just a China problem?

Not at all, and the researchers raising the alarm are careful to say so. Academics writing in response to a piece by China's ambassador to the UK point out that the gap between "open" as a marketing word and "open" as a technical standard exists across the whole industry.

Many Western labs also release model weights while keeping training data locked away. The argument for genuine openness, including shared standards that any country's AI must meet to claim the label, is one that applies everywhere.

For ordinary users, the practical takeaway is simple. When an AI tool is described as "open", it is worth asking: open how? Can anyone inspect the data it learned from? Can anyone audit who controls it? If the answers aren't clear, the word "open" is doing a lot of heavy lifting.

Common questions

Why does it matter whether AI training data is public?

If no one outside the lab can see what data an AI learned from, no one can check whether it contains biases, errors, or restricted information. Transparency about training data is one of the main ways researchers and regulators hold AI systems to account.

What is the model openness framework?

It is a standard for judging how genuinely open an AI system is, covering weights, training data, source code, and governance. A recent UN report endorsed it as a benchmark, meaning more international bodies are starting to use it when evaluating AI tools promoted for global use.