Musk: xAI Partly Used OpenAI Models to Train Grok

Elon Musk testified Thursday in a federal court in California that his company xAI partly used OpenAI models through a process known as distillation to train the Grok chatbot. The comment came during Musk’s trial against OpenAI, CEO Sam Altman and co-founder Greg Brockman.

When questioned about applying distillation techniques to OpenAI models, Musk replied “partly” and described the approach as a common industry practice.

Distillation trains a new system by querying an existing model through its public interface or an application programming interface and using the responses as learning signals. The method can allow developers to reproduce aspects of a model’s behavior at lower cost than building a system from scratch.

The practice has drawn scrutiny. Earlier this year, a rival AI developer accused several foreign teams of using fraudulent accounts to extract large volumes of responses from its chatbot to train competing systems. The White House warned of “industrial-scale” campaigns that use proxy accounts and jailbreaks to reproduce U.S. AI capabilities.

Legal experts have noted that the boundaries around distillation remain unclear. The technique is not explicitly illegal, but it can raise questions about whether it violates platform rules or API terms that govern how models may be queried and how outputs may be used for training.

xAI launched in July 2023 to develop its own models. Musk co-founded OpenAI in 2015 with Sam Altman, Greg Brockman, Ilya Sutskever, John Schulman and Wojciech Zaremba and left the organization in 2018. The trial examines whether OpenAI’s later shift toward a for-profit structure breached its original nonprofit charter and Musk’s expectations when he helped start the group.

OpenAI and xAI did not immediately respond to requests for comment.