Anthropic Claims Alibaba Misappropriated Claude's Capabilities

AI startup Anthropic claims that China's major technology company Alibaba improperly leveraged the capabilities of its large language model 'Claude.' The company alleges that Alibaba used a technique called 'model distillation' to use Claude's outputs as training data and enhance the performance of its own model.

Model distillation refers to a technology that efficiently trains another model (student model) by using the output results of a high-performance AI model (teacher model). While this technique is widely used in legitimate applications, using large volumes of another company's model outputs without permission in service terms to strengthen one's own model may constitute intellectual property infringement. Anthropic's concern is specifically about this unauthorized usage beyond the terms of service.

AI companies generally prohibit obtaining outputs through API (Application Programming Interface) access to their models and using those outputs to train competing models, as stated in their service agreements. Anthropic has similar provisions in place, and this claim points to a violation of those terms. Meanwhile, official comments or rebuttals from Alibaba have not been confirmed at this time.

The emergence of this issue is rooted in intensifying AI development competition. As companies seek to develop higher-performance models at lower costs, there is a growing incentive to exploit outputs from competing models as a 'loophole.' Particularly in the current environment where closed, high-performance models are widely deployed globally, their output data can effectively serve as a shortcut to development.

This case is not irrelevant to companies using AI services. When leveraging AI for their own data and operations, the importance of conducting due diligence to verify how the AI provider manages and uses data is underscored. The massive amounts of usage data collected by AI labs contain risks of misuse by competitors, and companies using these services must scrutinize service terms and data management policies.

Intellectual property disputes among AI companies are likely to surface further in the future. With global rules regarding the use of training data and outputs still in development, the conflict between Anthropic and Alibaba serves as an example that calls into question the governance framework of the entire industry. Going forward, attention will focus on legal determinations and developments in official statements from both companies.