Elon Musk has acknowledged under oath that xAI partly trained its Grok chatbot by distilling outputs from OpenAI's models, a tactic that breaches OpenAI's terms of service even as it has become common across the AI industry. The admission came during testimony in federal court in California tied to the long-running litigation between Musk and OpenAI chief executive Sam Altman.
Musk was asked directly whether xAI had used distillation, the process of asking a rival model questions, capturing its answers, and feeding those question-and-answer pairs back into a new model as training data. He conceded the company had done so in part, framing the technique as a normal step in how modern frontier labs build their systems.
"Elon yesterday, was in federal court in California in the lawsuit, and he was asked whether XAI use distillation on open AI models to train grok," Jaeden Schafer said on the podcast, walking through the day's testimony. The mechanic, he explained, amounts to "basically asking open AI model a question, getting an answer, and then feeding the answer, the question and the response back to grok to help train it."
Key facts
- 01Elon Musk testified in federal court in California that xAI partly used distillation on OpenAI models to train Grok.
- 02Distillation involves querying a rival model and feeding the question-and-answer pairs back into a new model as training data.
- 03The practice violates OpenAI's terms of service but is not illegal, and DeepSeek used the same method to train its models.
- 04Anthropic has lobbied the government to force companies to stop the practice, according to comments on the podcast.
The disclosure lands awkwardly for Musk, who has positioned xAI as a rival built from scratch and has publicly accused OpenAI of betraying its non-profit roots. It also sharpens a debate the industry has so far preferred to keep quiet: most leading labs are believed to lean on competitor outputs to accelerate training, even when the source provider explicitly bars the practice.
“He tried to kind of wave it off saying it's kind of general industry practice, which, to be fair, yeah, it is deep seek. That's how deep seek did it, like, actually, every AI model company does it.”— Jaeden Schafer
"He tried to kind of wave it off saying it's kind of general industry practice, which, to be fair, yeah, it is deep seek. That's how deep seek did it, like, actually, every AI model company does it," Schafer said. China's DeepSeek drew scrutiny earlier this year for the same approach, and OpenAI has previously suggested its outputs were used to seed competing systems.
Distillation is attractive because it compresses what would otherwise be a slow, expensive labelling process. By harvesting fluent answers from a stronger model, a smaller team can teach its own system the shape of a good response far faster than by paying human annotators. Schafer noted the trade-off: the technique speeds training "because you're telling it Look, when someone asks this type of question, this is the kind of response that we would like you to give," but caps the student model's ceiling at roughly the teacher's quality.
The legal exposure is narrower than the optics suggest. Violating OpenAI's terms of service is a contract issue, not a criminal one, and there is no US statute that bans training on another model's outputs. Anthropic has reportedly pushed Washington to change that, lobbying for rules that would force labs to stop distilling each other's models, according to comments on the podcast.
For Musk, the bigger risk may be reputational. Court testimony admitting that Grok was partly trained on OpenAI outputs hands Altman's lawyers a useful talking point and complicates xAI's pitch as an independent challenger. With weeks of hearings still to come, daily disclosures from the case look set to keep reframing how the AI industry's two highest-profile founders see each other — and how regulators see them both.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
