← Back to feed

On April 30, 2026, in a federal courtroom in Oakland, Elon Musk was asked a direct question by OpenAI attorney William Savitt: had xAI used distillation on OpenAI models to train Grok? After initially deflecting — calling distillation a general industry practice — Musk answered: “Partly.”

The word carried more weight than Musk appeared to intend. He is suing OpenAI for abandoning its nonprofit mission. The testimony established that his own company built its flagship AI by querying a competitor’s models and feeding those outputs back as training data.

What Distillation Is and Why It Matters

Model distillation involves prompting an existing AI model at scale, collecting its question-and-answer pairs, and using them to train a new model. Smaller or newer labs use the technique to compress frontier-level knowledge into their own systems at a fraction of the compute cost required to train from scratch.

The practice is not illegal under US law. But it typically violates the terms of service that OpenAI, Anthropic, and Google set for API access. All three companies have moved aggressively against what they describe as systematic distillation — primarily aimed at Chinese firms. OpenAI and Anthropic have both named DeepSeek and others in public complaints, and the Frontier Model Forum has been used to coordinate detection of suspicious mass-querying.

Musk’s testimony on April 30 was the first public under-oath admission from a US frontier lab founder that the practice occurred at a top American company.

The Ranking That Tells the Story

The distillation exchange was not the only consequential moment in Musk’s testimony. When asked about his earlier claim that xAI would soon surpass nearly every competitor except Google, he offered a markedly different picture. Under oath, he ranked the world’s leading AI providers as follows:

1. Anthropic — currently leading on capability and safety
2. OpenAI — close second
3. Google
4. Chinese open-source models — rapidly advancing
5. xAI — characterised as “much smaller,” a few hundred employees

That ranking sits uncomfortably against xAI’s public positioning. The company merged with X in 2025, was subsequently folded into SpaceX, and has consistently marketed Grok as a top-tier frontier model. Musk’s own testimony suggests it is not.

The Legal Irony

Musk’s lawsuit alleges that OpenAI’s shift from nonprofit to for-profit structure caused harm to the public and the AI industry. OpenAI’s legal team has responded by producing early emails showing Musk himself proposed a for-profit structure in 2017 before his 2018 departure from the board.

Judge Yvonne Gonzalez Rogers noted from the bench that it was “ironic” for Musk to raise AI safety concerns while building a company in the same field. The distillation admission sharpened that observation: xAI’s product was built, at least in part, by extracting value from the closed models Musk is suing to open.

OpenAI has not publicly responded to the admission or indicated whether it will pursue a terms-of-service claim. The legal exposure for xAI is likely limited — contract disputes, not criminal liability — but the reputational damage lands squarely in the narrative Musk built his case around.

The trial continues. Greg Brockman and AI safety researcher Stuart Russell were next on the witness list following Musk’s dismissal.