Musk Under Oath: Grok Was Trained on OpenAI Outputs

Elon Musk sued OpenAI in August 2024 on the theory that Sam Altman and Greg Brockman had defrauded him: he co-founded and bankrolled OpenAI as a nonprofit whose charter called for safe, open AI development, and they converted it into a for-profit structure designed, in his telling, to enrich insiders and Microsoft rather than humanity. Of the 26 original claims, two survived to trial: breach of charitable trust and unjust enrichment, the latter tied to Altman's escalating compensation as OpenAI's valuation approached $1 trillion. The liability phase opened April 28, 2026, before Judge Yvonne Gonzalez Rogers in case 4:24-cv-04722-YGR in Oakland's Northern District. On April 30, four days in, Musk took the stand and, under cross-examination by OpenAI lead attorney William Savitt, acknowledged that xAI had "partly" used knowledge distillation from OpenAI's models while building Grok.

The Grok 4 model card, published by xAI in August 2025, describes training on public internet data and content from X, the platform Musk also controls. Nothing in it mentions outputs from OpenAI's GPT series or any competing model. The original Grok-1 open-source release, in March 2024, characterized that model as trained "from scratch." When Savitt pressed him to define distillation from the stand, Musk answered: "It means to use one AI model to train another AI model," then added, "It is standard practice to use other AIs to validate your AI."

Knowledge distillation through commercial APIs works by submitting prompts to a target model, collecting its outputs, and using those question-answer pairs to fine-tune a smaller model. OpenAI's terms of service have prohibited using its outputs "to develop models that compete with OpenAI" since at least early 2023, before xAI was incorporated and before any Grok model shipped. Whether xAI accessed those outputs through a paid enterprise account, a standard consumer subscription, or some other channel was not established in published trial accounts; nor were the specific Grok releases that drew on distillation.

On February 12, 2026, OpenAI submitted a memo to the U.S. House Select Committee on China accusing DeepSeek of "ongoing efforts to free-ride on the capabilities developed by OpenAI and other US frontier labs," citing evidence of DeepSeek employees circumventing API controls to harvest outputs. That memo produced no lawsuit, no takedown, and no formal legal filing. The accusation existed as a congressional notification and a sequence of public statements, nothing more. Two and a half months later, the man who had filed that position in a government hearing room confirmed from the witness stand that his own company had used the same technique.

Musk specified none of it on the stand: which OpenAI models, which Grok versions, when the distillation ran. "Partly" could mean a minor auxiliary fine-tuning set or a substantial share of the training signal. xAI did not respond to requests for comment on the testimony. Whether the admission damages Musk's charitable-trust case is a question Savitt left open on April 30; the answer will arrive in subsequent testimony or the remedies phase, if Musk prevails on liability.