Musk Confirms xAI Used OpenAI Models to Train Grok
SAN FRANCISCO — Elon Musk testified Thursday that his AI startup xAI used OpenAI’s models to train its Grok chatbot through a technique known as model distillation, according to The Verge.
The admission came during the ongoing legal battle between Musk and OpenAI, the company he co-founded before departing and later suing. Model distillation is a widely used practice in the AI industry in which a larger, more capable model effectively acts as a teacher, transferring knowledge to a smaller model to improve its performance.
The admission places the practice at the center of a federal court proceeding. Musk has been among OpenAI’s most vocal critics, accusing the company of straying from its original open-source mission.
The testimony comes as Musk has pursued legal action against OpenAI on multiple fronts, even as his own company appears to have relied on OpenAI’s technology to develop its competing product.
What Is Model Distillation?
Model distillation is a common technique in machine learning where the outputs of a large, computationally expensive model are used to train a smaller, more efficient one. The smaller “student” model learns to approximate the behavior of the larger “teacher” model, often achieving comparable performance at a fraction of the computational cost.
The practice is widespread across the AI industry, but it exists in a legal gray area. Most major AI providers prohibit using their model outputs to train competing systems in their terms of service. Whether such contractual restrictions are enforceable — and whether distillation constitutes intellectual property infringement — remains an open legal question.
Industry Implications
Musk’s courtroom admission could have ramifications well beyond the OpenAI-xAI dispute. If the court addresses the legality and boundaries of model distillation, the ruling could set precedent for how AI companies protect their models and what competitors can lawfully do with publicly available AI systems.
Several AI providers, including OpenAI, Google, and Anthropic, include provisions in their terms of service that restrict using model outputs for training purposes. Enforcement of these provisions, however, has been inconsistent, and the industry lacks clear legal frameworks governing the practice.
The case is being closely watched by AI researchers, legal scholars, and competing AI labs as one of the first major federal proceedings to directly address intellectual property questions surrounding large language models and the data used to build them.
The trial is ongoing in federal court in California. Neither xAI nor OpenAI immediately responded to requests for comment beyond their existing court filings.
