In a federal courtroom in California on Thursday, Elon Musk testified that his own AI startup, xAI, has used OpenAI’s models to improve its own. The confession cuts to the heart of a practice that frontier AI labs have started fighting hard to prevent.
The technique in question is called model distillation. It’s a common industry practice by which one larger AI model acts as a teacher to pass on knowledge to a smaller model. In this case, xAI appears to have used OpenAI’s systems to train Grok, the company’s flagship chatbot. Distillation is a hot topic as frontier labs try to prevent smaller competitors from copying their models.
The stakes here are obvious. If you can take a large, expensive model built by one company and use it to train a competing product, the entire moat around AI development gets eroded. OpenAI has spent billions building ChatGPT and its underlying systems. The idea that a rival startup could simply feed those same systems into their own training process and emerge with a competitive alternative is exactly what the industry wanted to prevent.
Musk’s testimony is significant because it confirms something the AI world has suspected for months. Model distillation isn’t theoretical anymore; it’s happening. And one of the industry’s most prominent figures just admitted to doing it in court.
What makes this different from past instances of competitive tension is the admission itself. In earlier disputes over AI model copying, companies claimed they hadn’t used proprietary systems. This time, the founder stood up in court and said yes, we did that. We took their model and used it to train ours. The practice itself wasn’t secret, but the scale and directness of the approach is now documented in legal proceedings.
The courtroom battle signals how seriously OpenAI takes model security. As the undisputed leader in the generative AI space, the company has every incentive to lock down its technology. But the fact that model distillation works at all suggests that no system is fully protected. If you have access to a model’s outputs, you can use those outputs to train something else. The teacher-student dynamic means the student gets smarter without needing access to the teacher’s internals.
For the broader AI industry, this is a moment of reckoning. The same model distillation technique that xAI used is available to any team with computational resources and API access. That changes how we think about competitive advantages in AI development. Size and resources still matter. But so does the ability to be clever about training and engineering.
The real question now is whether Musk’s testimony will change how these disputes get settled. Legal precedent matters less in AI than it does in traditional tech because the technology moves so fast. By the time courts rule on model distillation, the entire landscape could shift. What we’re watching is the frontier labs trying to draw lines around what’s acceptable. Musk just admitted xAI crossed one of them.