Europe should let the market judge China’s open AI models

The case for Europe embracing China’s open AI models is more tempting now than it has ever been, for two reasons that have sharpened in recent months. 

The first is Washington. The United States has tightened its grip on advanced AI – restricting the chips that train the best systems, and this summer briefly pulling even its own most capable models from sale – a reminder that access granted is access that can be withdrawn. 

The second is that China’s models have moved from trailing to matching the frontier of US AI models. In fact, when Moonshot unveiled Kimi K3 at July’s World Artificial Intelligence Conference in Shanghai – the largest freely downloadable model yet built, ranked by some independent tests alongside the best American systems – it knocked more than one percentage point off the Nasdaq in a single day. 

Put the two together and European policymakers seem to be toying with the idea of turning towards China’s free models as an insurance for an unreliable partner, an insurance which also comes at a cheaper cost in terms of Europe’s potential bill for its AI dependence.

The temptation is certainly not foolish. Excellent and cheap Chinese models could reduce our dependence on an unpredictable Washington. A model you run on your own servers is, in one narrow sense, more sovereign than one rented from California. 

It would be best to let market forces choose which models take root in Europe. Yet there is a difference between letting the market decide and states steering the decision. European firms are already free to download these models and run them on their own servers, and many do. But Europe should not turn a commercial choice into official policy by steering public money, government contracts and its own sovereignty story toward China’s technology. Why?

First, “free” is a business strategy, and business strategies get reversed. China’s open models are not a gift to the world; they are a way to win users cheaply now and sell them services later. Most of the labs behind them make little or no money and are kept going by patient backers – hedge-fund parents, tech giants, state-linked investors – who are financing a land-grab, not a public good. The moment that calculation changes, openness is the first thing to go.

It is already happening: Alibaba, once among the loudest champions of giving its models away, now keeps its most powerful version behind a paywall. The honest version of the risk is not that today’s downloads vanish – once released, they stay released – but that the stream of upgrades and security fixes dries up, leaving whoever built on them stranded on a model that quietly ages, with no supported path forward.

Second, leaning on Chinese models could backfire precisely as the two AI superpowers decouple. The Americans keep their best systems closed, the Chinese give theirs away, and Europe sits in the thin space between – a space that is narrowing. Washington has shown it will restrict access when it chooses, and its next target may be the Chinese models themselves. 

Many are widely believed to have been trained partly on the outputs of American systems – a shortcut known as distillation – and the US is weighing sanctions to stop it. If those come, the effects will impact Europe in two ways: the Chinese models could lose the cheap route to improvement they have relied on and fall behind; European firms that built on models tainted by restricted American technology could find themselves caught by US enforcement. A quiet commercial choice is easy to unwind; a public policy bet is not.

Third, an official embrace of Chinese models could smother Europe’s own. Sovereign models such as France’s Mistral do not yet match the frontier, and they never will without European companies and governments choosing to buy them, train on them and push them up the curve. Brussels has finally put money and political capital behind that goal, from its new sovereignty and chip plans to shared public computing power. All of it assumes European buyers will back European models. Signalling that the sovereign answer already exists, for free, from Hangzhou, quietly starves them of the demand they need to survive.

Fourth, you cannot govern what you cannot see inside. These models are “open” only in a narrow sense: you can download and run them, but you cannot inspect the data they were trained on or the instructions baked into them. That gap is not neutral. Independent testing has found Chinese models quietly avoiding politically sensitive subjects and shading their answers on anything touching China, and behaving differently in Chinese than in English. 

Europe’s whole comparative advantage is that it governs technology by clear rules – on transparency, privacy and fundamental rights. Wiring public services to a model whose inner workings no outside party can audit sits badly with that ambition and invites problems, from compliance to security, that tend to surface only later.

The answer is not to ban Chinese AI models, but nor should Europe promote them. European firms weighing cost, licensing, and where their data lives, should be free to choose Chinese models – but they should not be pushed into doing so. 

European policy makers’ task is not to pick a foreign winner. It is to keep the market open, price the risks honestly, and spend its scarce political capital building strength at home. Let the market judge China’s models on their merits. Brussels should keep its thumb off the scale.

Alicia García Herrero is a Senior Fellow at Bruegel. She specialises in emerging markets, with a particular focus on Asia and Asia-European Union relations. She is also Chief Economist for Asia Pacific at French investment bank Natixis.