U.S.-based artificial intelligence firm Anthropic has accused three Chinese AI laboratories of conducting coordinated, industrial-scale distillation campaigns against its flagship model Claude, intensifying scrutiny over intellectual property protections in the global race for advanced AI systems.
In a public statement, Anthropic said it identified large-scale activity linked to DeepSeek, Moonshot AI and MiniMax that allegedly involved the creation of more than 24,000 fraudulent accounts. These accounts generated over 16 million exchanges with Claude in what the company described as a coordinated effort to extract model capabilities and improve competing systems.
According to Anthropic, the activity went beyond ordinary commercial use. The firm said traffic patterns were synchronized across accounts, shared payment methods were detected and query timing suggested deliberate distribution of load to avoid detection. In certain cases, prompts allegedly attempted to extract detailed reasoning explanations from Claude, which Anthropic claims could be used to reconstruct model behaviour for training rival models.
Model distillation is a widely used technique in artificial intelligence development. It typically involves using a larger and more capable system to generate outputs that help train a smaller or more efficient model. Companies rely on this approach to reduce costs, optimize deployment and improve performance in specific tasks. Anthropic, however, argues that large-scale extraction through coordinated accounts and policy circumvention violates its terms of service and amounts to intellectual property misuse.
The company also framed the issue in strategic terms, warning that illicit distillation could weaken safeguards built into advanced models and potentially allow capabilities to be integrated into sensitive applications, including military or surveillance systems. Such language places the dispute within the broader geopolitical contest between the United States and China over control of advanced technologies.
The allegations have sparked debate within the technology community. Some analysts question where the boundary lies between aggressive competitive behavior and contractual violation, particularly when model outputs are accessible through paid APIs. Others note that leading AI firms have themselves relied on vast quantities of publicly available internet data, licensed datasets and scraped content during pretraining, raising complex questions about ownership and reuse of machine generated outputs.
The episode also reopens a deeper structural debate about the foundations of generative AI. Much of the modern AI ecosystem has been built on large-scale data aggregation across digital platforms, and courts in multiple jurisdictions are currently examining whether certain training practices constitute fair use or copyright infringement. In that context, critics argue that disputes over distillation reflect a broader unresolved tension within the industry itself. If intelligence is derived from data that originated in the public domain or from copyrighted sources under legal dispute, then questions of ownership, replication and competitive advantage become structurally complex.
As foundation models grow more powerful and increasingly expensive to train, the incentive to approximate their capabilities through large-scale querying will likely rise. The current dispute therefore reflects not only a bilateral corporate conflict, but a deeper systemic challenge over intellectual property, competitive moats and the economics of large language models in an API-driven global marketplace.
