Elon Musk confirms xAI used OpenAI’s models to train Grok

Summary

In his federal court testimony in the Musk v. Altman case, Elon Musk admitted that OpenAI models were distilled to develop Grok. Musk characterized this as a 'standard practice,' dismissing the issue by stating that 'generally, all AI companies do the same thing.' By limiting his answer with the word 'partially,' Musk avoided giving a clear yes or no.

Why it matters

This admission came during an active lawsuit. It transformed the debate over model distillation from a hypothetical sector concern into a confirmed competitive practice between the two largest AI labs. It also highlighted a legal gray area. While distillation may be legitimate within a company, when done between competitors, it could constitute a breach of terms of service or intellectual property theft. Regulators and courts are now forced to carefully evaluate this distinction.

Highlights

• Musk's 'partially' response and his defense that 'all companies do this' prevent a direct admission regarding scale or method.
• The report notes that OpenAI, Anthropic, and Google have accused competitors, particularly Chinese labs, of illegal distillation.
• Anthropic's own framework: Distillation is legitimate within a company. However, it can also be used for 'illegal purposes' to acquire capabilities 'very quickly.'

Critical take

Musk's 'partially' answer and 'standard practice' framing are classic courtroom evasions. This statement does not clarify how much of Grok's capabilities were obtained through distillation versus developed through independent training. While The Verge's report is based on good sources, it largely repeats Musk's evasion. It fails to question whether xAI's usage violated OpenAI's terms of service or whether the 'partially' qualifier itself constitutes a legal loophole.