As the global artificial intelligence race intensifies, “AI distillation” has emerged as a new focal point in technological competition between the United States and China. Originally a model optimisation technique, it has recently become entangled in disputes over intellectual property and alleged technological replication. The United States has accused some Chinese firms of using distillation methods to “replicate” the capabilities of advanced AI models, while China denies the claims and insists its progress is driven by independent innovation.
At its core, AI distillation involves using a large “teacher model” to generate outputs, which are then used to train a smaller “student model.” This allows the smaller system to achieve performance close to that of larger models at significantly lower cost. While the technique reduces computing requirements and accelerates AI adoption, it also makes it easier for advanced model capabilities to be transferred and disseminated.
The US government has recently expanded its regulatory focus beyond chip exports to include algorithms and data, framing AI competition as a form of “arms race.” Donald Trump is also expected to discuss AI leadership issues with Xi Jinping at an upcoming meeting, highlighting the growing centrality of technology in bilateral relations.
In the industry, some companies have acknowledged using outputs from other AI systems for distillation, raising ethical concerns over “innovation” versus “free riding.” However, experts note that the practice is widespread and difficult to clearly define within existing boundaries.
As semiconductor restrictions tighten, Chinese firms are increasingly turning to distillation to improve efficiency and reduce reliance on high-end hardware, shifting competition from “computing power advantage” toward “efficiency and algorithms.” Nevertheless, the actual impact of distillation on model performance remains difficult to quantify, creating new challenges for regulation and intellectual property frameworks.
Commentary
From a technical standpoint, AI distillation is closer to “knowledge reuse” than “technology theft.” In modern machine learning systems, using model outputs as training data is already a common and established practice. If every instance of learning from another model’s output is immediately elevated into a matter of national security or intellectual property theft, then the issue risks being overly politicised.
Rather than treating distillation as a threat, it may be more accurate to see it as part of an irreversible process of technological democratisation. The real competition is no longer about who can monopolise the most powerful model, but who can use, adapt, and integrate AI systems more effectively.