A storm is brewing in the quiet corridors of British intellectual property law, and its epicentre is a high-stakes clash between Anthropic, the San Francisco-based AI safety lab, and Alibaba, the Chinese e-commerce and cloud computing behemoth. Last week, Anthropic formally accused Alibaba of stealing its proprietary AI model. The accusation is not just a corporate squabble. It is a test case for the United Kingdom, where the government has been scrambling to position itself as the global hub for AI innovation without sacrificing the rule of law.
Anthropic, a company born from the exodus of disillusioned OpenAI employees, has built its reputation on responsible AI development. Its flagship model, Claude, is trained with a constitution of ethical constraints designed to prevent harmful outputs. The company alleges that Alibaba’s latest large language model, Qwen3, shares uncanny architectural similarities with Claude. More damningly, Anthropic claims that Qwen3 exhibits identical responses to a set of proprietary prompts that were deliberately engineered to act as a watermark, a tell-tale sign of copying.
Alibaba, for its part, has dismissed the allegations as “baseless” and “a desperate attempt to slow down Chinese technological progress.” In a statement, the company insisted that Qwen3 was built from scratch using publicly available research and its own innovations in sparsity and efficiency. But the evidence presented by Anthropic’s legal team is compelling. They have provided a side-by-side comparison of model outputs, including responses to nonsensical inputs that an independent model would never replicate verbatim.
This is where British intellectual property law enters the fray. The case could be heard in the UK courts because Anthropic has a significant presence in London and may argue that Alibaba’s model is being distributed through cloud services accessed by British users. The UK’s intellectual property framework, which has traditionally focused on patents, copyrights, and trade secrets, is now facing the slippery concept of “model theft.” How do you patent a neural network’s behaviour? How do you copyright a probabilistic output? The law, drafted long before the era of generative AI, is being forced to evolve.
The implications are profound. If the court rules in favour of Anthropic, it could set a precedent that the United States and Europe will watch closely. It would mean that the architecture and even the behaviour of an AI model can be protected under existing intellectual property laws. That would give a huge advantage to first movers, entrenching the market power of companies like OpenAI, Google, and Anthropic. It would also open the door to a wave of litigation against smaller AI labs in the Global South that may have inadvertently trained on similar data.
Conversely, if Alibaba wins, it could gut the notion of proprietary AI models. It would signal that any model’s weight configuration, once released into the wild, is fair game for replication and optimisation. This would accelerate the commoditisation of AI, benefiting open-source communities and lowering the barrier to entry for startups. But it would also undermine the incentive for companies to invest billions in safety research, which Anthropic argues is a cornerstone of responsible AI development.
There is a darker angle to this story, one that Julian Vane, a Silicon Valley expat now based in London, captures with his usual prescience. Vane, who left Google in 2019 to start an ethics consulting firm, has been warning that the current intellectual property regime is not fit for purpose. “We are relying on eighteenth-century concepts to govern twenty-first-century algorithms,” he told me. “The idea that you can ‘steal’ a model is like saying you can steal a snowflake. It melts the moment you touch it. What you are really stealing is the training process, the data, the context. And that is where the law needs to look.”
Vane’s point is that model theft is not like stealing a patent. A patent is a fixed design. A model is a dynamic, evolving entity that reflects its training data. The real theft, if there is any, is in the misuse of proprietary data to extract competitive advantage. But proving that requires access to Alibaba’s training pipeline, which is a black box. The British court will have to decide whether the evidence of identical outputs is sufficient to infer copying, or whether it is mere coincidence in the noisy world of deep learning.
As the case proceeds, the British government will be watching with a mixture of pride and anxiety. It has invested heavily in creating a “pro-innovation” regulatory environment, but a high-profile case like this could either cement the UK’s reputation as a jurisdiction that can handle cutting-edge disputes or reveal the cracks in its legal foundation. The outcome will echo far beyond the courtroom. It will shape the future of AI globalisation, the balance of power between Silicon Valley and Chinese tech giants, and the extent to which we can trust the models that will soon govern our healthcare, our finance, and our democracy.
For now, the world waits. The algorithms are watching too.
