Musk Admits xAI Distilled OpenAI Models for Grok

Added: 2026-05-11

[00:00:05]Welcome back to the Awesome Agents podcast. I'm Alex, and today we are talking about a courtroom moment that honestly I did not expect to see in 2026.

[00:00:16]Yeah, this one landed with a thud. Elon Musk under oath admitted that xAI used OpenAI's models to train Grok.

[00:00:24]Under oath in federal court. And the word he used was partly. Right, partly.

[00:00:31]Which is doing a lot of work in that sentence. Let me set the scene first because the context matters a lot here.

[00:00:38]Please, yeah. So, this is the Musk v.

[00:00:40]Altman trial. Musk is suing OpenAI, Sam Altman, and Greg Brockman claiming they abandoned the nonprofit mission that supposedly guided the company's founding. So, he's the plaintiff. He's the one who brought this case. Exactly.

[00:00:54]And then under cross-examination by OpenAI's lawyers, he gets asked whether xAI used distillation techniques on OpenAI models. He first calls it a general practice among AI companies and then just confirms it. Okay, so before we go further, let's make sure everyone knows what distillation actually means here. Because it sounds technical. It's not that complicated. You take a model someone else launched, you query it through its public API a massive number of times, you collect all those inputs and outputs, and you use that data to train your own model. So, you're basically teaching your model to mimic theirs. That's a fair way to put it. And the key thing is that it's distinct from just using outputs for research or testing. We're talking about feeding those pairs directly into a training pipeline. And every major lab has terms of service that explicitly ban this, right? All of them. OpenAI says you can't use their outputs to train competing models. Anthropic has the same clause. And here's the part that really got me. xAI's own Grok developer agreement has that same prohibition. So, Musk's company bans the exact thing Musk admitted to doing. Under oath, yes.

[00:02:02]Okay, and now we have to talk about the elephant in the room, which is the whole China narrative.

[00:02:09]This is where it gets genuinely uncomfortable. For most of 2026, US AI labs have been pushing a very clear story. Chinese companies are stealing American AI capabilities through distillation. It's a national security threat. The White House has said it.

[00:02:24]OpenAI accused DeepSeek directly. And Anthropic filed a formal complaint against DeepSeek, MiniMax, and Moonshot AI. Right? Like 24,000 fraudulent accounts scraping Claude at scale.

[00:02:35]24,000 accounts, 16 million exchanges, Anthropic called it theft. And now Musk, in a case where OpenAI is a defendant, has confirmed that xAI did the same thing to OpenAI.

[00:02:48]That's a hard position to sustain. You can't call it theft when a foreign competitor does it and standard practice when you do it. Musk tried to frame it as validation. Like they were checking if Grok's outputs were comparable to other frontier models. Is that legitimate? Validation is legitimate.

[00:03:03]Running a few thousand queries to benchmark your model against competitors, completely normal. But when you're collecting millions of pairs and feeding them into training runs, at some point the line between validation and extraction gets pretty blurry.

[00:03:16]And we don't know how much of Grok's training data came from this. He just said partly. Right, which is strategically vague. It's enough to avoid perjury, but not enough to tell us whether it was 0.1% or 20% of the data.

[00:03:29]What's interesting to me is that even with this apparent advantage, Grok still isn't at the top of the benchmarks.

[00:03:34]That's a good observation. Grok 4.20 scored 49 on the artificial analysis intelligence index. GPT 5.4 and Gemini 3.1 Pro are both ahead of it. So, whatever distillation contributed, it didn't close the capability gap. Which maybe tells you something about the limits of the technique. It does. You can copy behavior patterns, but you can't easily distill whatever it is that makes a frontier model strong at reasoning on truly novel problems. What does this mean practically for developers who are building on these APIs? The risk is that the independence you assume between two competing systems might not be real. If GPT and Grok share training lineage through distillation, benchmarks comparing them are partly circular. You're not measuring two independent approaches, you're measuring how well one copied the other. That's a real issue for anyone trying to pick the right model for a critical application.

[00:04:24]And the legal picture is still genuinely unsettled. Violating terms of service is a civil matter. Courts haven't ruled on whether distillation counts as theft of model capabilities or just use of a public API. The IP law doesn't cleanly cover it yet. Judge Gonzalez Rogers has already been pretty skeptical about Musk's framing across the trial. She said she suspects there's plenty of people who don't want to put the future of humanity in Mr. Musk's hands. Week two starts May 5th, so there's more to come. This case might end up reshaping how these terms of service get enforced, or at least whether anyone takes them seriously. The honest read is probably that Musk is right that other companies have done this, too. The industry has just agreed not to say so publicly. What changed is that a CEO said it under oath in federal court with OpenAI sitting across the aisle.

[00:05:10]That's what changes the legal record, even if it doesn't change the technical reality on the ground. Exactly. We'll link to the full coverage in the show notes, including the Anthropic distillation complaint and the trial reporting from MIT Tech Review and TechCrunch.

[00:05:25]Thanks for listening to the Awesome Agents podcast. We'll be back with more as week two of the trial wraps up. See you then.