Elon Musk has spent months suing OpenAI, alleging the company betrayed its nonprofit mission and caused him personal harm. Now he's acknowledged that his own AI models at xAI were partially trained using OpenAI's technology.
The admission creates obvious tension with the legal position xAI has taken. It doesn't automatically constitute wrongdoing - training relationships between AI companies are complicated and common - but it does chip away at the cleaner narrative Musk has built around xAI as the principled alternative to ChatGPT.
The specific technology involved matters a lot here. AI companies build new models in a few ways: training on raw internet data, using proprietary datasets, or through a process called "knowledge distillation" - where a new model learns by mimicking the outputs of an existing, more capable model. If xAI's Grok models were distilled from OpenAI's outputs, that would likely violate OpenAI's terms of service, which explicitly prohibit using their models to train competing systems.
Musk didn't specify which method was involved, and xAI hasn't elaborated publicly.
Who Else Has Done This
xAI wouldn't be alone if the answer is model distillation. Meta's early LLaMA models, Mistral's first releases, and a long list of smaller open-source projects have faced credible questions about whether they trained on ChatGPT outputs. The practice became common enough in 2023 that researchers gave it a name - "ChatGPT distillation" - and published papers documenting how widely it had spread through the open-source AI community.
Most companies quietly moved on. Musk is in a different position because he's actively litigating against OpenAI.
The lawsuit has been working through courts without a decisive outcome. OpenAI's lawyers will likely use this admission to argue that Musk's complaints about the company's competitive behavior look different when his own company was relying on OpenAI's work to get started. Whether that affects the legal outcome is a separate question - breach of mission and training practices are different legal claims - but it hands OpenAI a useful counterargument in public.