xAI Admits Distilling OpenAI Models in Court , and Changes the IP Stakes for Everyone

2. xAI Admits Distilling OpenAI Models in Court , and Changes the IP Stakes for Everyone

In week one of the Musk v. OpenAI trial in San Francisco, Elon Musk took the stand and made three distinct claims: that Sam Altman and Greg Brockman deceived him into funding OpenAI's early operations, that AI poses existential risk to humanity, and , most consequentially for the industry , that xAI has distilled OpenAI's models to train its own systems. The distillation admission came under examination and was not a voluntary disclosure. The trial, which began late April 2026, is the most high-profile AI legal proceeding to date.

The distillation admission reframes what was already a contentious lawsuit. OpenAI's terms of service explicitly prohibit using its model outputs to train competing systems. If xAI distilled Grok on GPT-family outputs, Musk's team has handed opposing counsel a concrete contractual violation to argue alongside the broader fiduciary claims. More broadly, this is the first time a major AI lab has confirmed in open court that it trained on a competitor's model outputs. That sets a precedent every legal team at Anthropic, Google DeepMind, Meta AI, and Mistral is now reading closely. The question of whether distillation constitutes IP infringement has been theoretical. It is no longer theoretical.

The pattern worth watching: distillation has been an open industry secret. Smaller labs and open-source projects have trained on GPT-4 outputs for years, and OpenAI has rarely pursued enforcement. A court finding against xAI on this specific point would give OpenAI, and potentially other frontier labs, a tested legal instrument to restrict the practice industry-wide. That changes the cost calculus for any team currently using frontier model outputs as training signal.

Source: MIT Technology Review