Nadella CEO Publicly Criticizes 'Distillation Bans' by OpenAI and Others

Satya Nadella, CEO of Microsoft, has publicly raised concerns about AI labs including OpenAI and Anthropic. He pointed out the contradiction in these companies' approach: they prohibit 'distillation'—a technique where the output of one model is used to train another model—in their terms of service, while simultaneously utilizing widely available public data from the internet for training their own models. He termed this asymmetric structure the 'reverse information paradox.'

'Distillation,' as mentioned here, refers to a technique of using the output results of large-scale AI models as training data to develop smaller or alternative models. Both OpenAI and Anthropic prohibit this method in their terms of service, preventing their own models from being used as 'learning material' by other companies. Meanwhile, these companies have drawn upon the concept of fair use as a basis for large-scale collection of various publicly available texts and images on the web as training data for their own model development. Furthermore, Nadella also noted that these companies continuously learn from interactions with customers.

The crux of Nadella's argument is that 'companies should be able to control their own learning infrastructure.' He argued that against the asymmetric stance of AI labs—freely utilizing others' data while strictly restricting the use of their own data—there is a need for an environment where customer enterprises can autonomously manage their own data and learning processes. It should be noted that Microsoft itself is in the position of providing such 'self-managed learning infrastructure' as a cloud service.

The background to this statement includes the fact that discussions about data usage rules are becoming increasingly active across the AI industry. Lawsuits regarding the legality of training data for large language models and copyright infringement are mounting worldwide, and 'whose data can be used, under what conditions' has become a common issue facing the industry. In this context, the view that major AI labs are operating under asymmetric rules favorable to themselves is already being shared among some researchers and competitors.

Nadella's statement is not merely a criticism and cannot be separated from business context. Microsoft, through its cloud platforms including Azure, provides an environment where companies can independently customize and train AI models using their own data. The act of advocating for 'self-management of learning infrastructure' can, as a result, amount to promoting Microsoft's suite of services—a structure that should be kept in mind when evaluating the statement.

The question of asymmetry surrounding training data for AI models can be viewed as a discussion that may influence future industry standards and rule formation. How major labs will explain the consistency between the restrictions they impose in their terms of service and their own learning practices remains a point that will continue to be questioned from perspectives of transparency and trustworthiness. As companies incorporate AI into their own operations, the perspective of how to ensure control over the handling of their own data and their learning processes is positioned as increasingly important going forward.