The man suing OpenAI for betraying its nonprofit roots just admitted under oath that his own AI company did exactly what he's accusing them of protecting against.
The Summary
- Elon Musk testified that xAI trained its Grok model using OpenAI's models, a practice called "distillation" that lets smaller labs copy capabilities from frontier models without doing the original research.
- Musk defended the practice as standard across AI labs, even as OpenAI and other frontier companies are fighting to prevent exactly this kind of competitive copying.
- The admission came during testimony in ongoing litigation, highlighting the tension between open practices Musk claims to support and the closed moats every serious AI lab is building.
The Signal
Distillation works like this: you query a powerful model thousands of times, collect its outputs, then use those responses to train a smaller, cheaper model that mimics the behavior without understanding how the original was built. It's pattern matching, not innovation. You get 80% of the capability at 20% of the compute cost.
Frontier labs hate this. They spend hundreds of millions on compute clusters, recruit the best researchers, run training runs that cost eight figures. Then someone downloads their API, runs 100,000 queries, and trains a knockoff for the price of a used Honda. OpenAI, Anthropic, and Google have all updated their terms of service in the past year to explicitly ban using their outputs for training competing models.
"Musk argued it's standard practice for AI labs to use their competitors' models."
But standard practice and legal practice aren't the same thing. Here's what makes this testimony remarkable:
- Musk is currently suing OpenAI, claiming they betrayed their nonprofit mission by going closed-source
- He's positioned xAI and Grok as the "anti-OpenAI," promising more openness and less corporate capture
- He just admitted under oath that xAI used the closed models he's suing about to train his own competing product
The irony is thick enough to stop a bullet. Musk's legal argument against OpenAI rests partly on them restricting access to models that should benefit humanity. His business strategy at xAI apparently involves using those same restricted models as training data. Both things can't be true at the same time.
The timing matters. This isn't ancient history. xAI launched in 2023. Grok shipped in late 2023. The distillation Musk described would have happened recently, when terms of service explicitly prohibited this use. Either xAI violated OpenAI's ToS, or Musk is describing something that happened before those restrictions went into effect, which would make Grok's training run older and less impressive than xAI has implied.
The Implication
Every frontier lab is now dealing with the same problem: how do you publish research, provide API access, and build an ecosystem without giving competitors a free blueprint? The answer so far has been lawyering, not technology. Terms of service get longer. Licensing gets more restrictive. The "open" in OpenAI becomes a nostalgic brand name, not a business model.
If Musk's description of distillation as standard practice is accurate, we're heading toward a world where the value isn't in the model weights or the architecture. It's in the training data, the compute infrastructure, and the legal moats you build around both. The agents economy gets built by whoever can afford the electricity bill and the lawyers, not whoever has the best ideas. Watch whether OpenAI updates their lawsuit strategy based on this testimony. And watch whether any other AI lab CEOs suddenly get very careful about what they say under oath.