1. Headline & intro
Elon Musk has finally confirmed what many inside the AI world assumed but could never quite prove: the frontier labs are quietly training on each other. His admission that xAI used OpenAI models to help build Grok isn’t just a spicy moment in a California courtroom — it blows open a debate about who really owns AI capabilities, and how far competitive copying can go before regulators and courts step in.
In this piece, we’ll unpack what Musk actually revealed, why model "distillation" has the biggest players so nervous, how this intersects with regulation and geopolitics, and what it means for developers, startups and users in Europe and beyond.
2. The news in brief
According to reporting from TechCrunch, Elon Musk testified in a California federal court that his company xAI has used distillation techniques involving OpenAI models to train its own chatbot, Grok. Distillation refers to systematically querying a powerful model and using its responses as training data for another model, effectively transferring much of the original model’s behaviour.
Musk was testifying in his lawsuit against OpenAI, CEO Sam Altman and co‑founder Greg Brockman, where he alleges they abandoned OpenAI’s original non‑profit mission. On the stand, when asked if xAI had distilled OpenAI models, he said this was a general industry practice and acknowledged that xAI had done so "partly".
TechCrunch notes that OpenAI, Anthropic and Google have already launched an initiative under the Frontier Model Forum to coordinate defences against large‑scale distillation attempts, especially from Chinese actors. During the same testimony, Musk also ranked Anthropic as the current global leader in AI, followed by OpenAI, Google and Chinese open‑source models, putting xAI well behind with only a few hundred staff.
3. Why this matters
Musk’s courtroom comment matters for three reasons: it confirms an open secret, exposes a strategic contradiction, and sharpens looming legal battles.
1. Confirmation of an open secret
Insiders have long suspected that major labs query each other’s public APIs and chatbots to accelerate their own training. Until now, most public discussion focused on Chinese groups distilling US models into cheaper open‑weight systems. Musk has now effectively said: yes, Western labs do it to each other too. That will harden attitudes at OpenAI, Anthropic and Google about how aggressively they should lock down their APIs, rate‑limit mass queries and watermark outputs.
2. Strategic contradiction for the giants
Frontier labs spent years arguing that scraping the open web — including copyrighted material — is a necessary sacrifice for progress. Now that their models are being scraped via distillation, they suddenly talk about "theft" and "unfair" free‑riding. Musk’s admission underlines the hypocrisy across the industry: everybody wants open access when they’re small and protectionism once they’re big.
3. Legal and commercial risk
Distillation lives in a grey area. As TechCrunch points out, it’s not clearly illegal in itself; the main issue is breaching terms of service and potential claims around trade secrets. But ToS violations scale poorly as a deterrent when the actors are well‑funded labs, sometimes in other jurisdictions. We are heading toward test cases where courts must decide whether a capability learned from another model’s outputs can be owned or monopolised at all.
The near‑term losers are likely mid‑tier competitors and open‑source communities who will face tightened access to powerful APIs. The winners: whichever giants best combine technical defences, legal muscle and lobbying to shape the rules around what counts as "improper" distillation.
4. The bigger picture
Musk’s remarks sit at the intersection of three broader trends in AI.
1. The shift from data scraping to model scraping
First, we’re moving from a world where the core asset was scraped web data to one where the key asset is model behaviour. Distillation is essentially scraping cognition: using an existing model as a generator of high‑quality synthetic data. That raises different legal questions than traditional copyright disputes. Copyright law was designed for books and images; it has almost nothing to say about training a new system on the answers of another.
2. From open science rhetoric to defensive alliances
Second, the Frontier Model Forum’s anti‑distillation efforts show how quickly "open research" rhetoric evaporates once incumbents feel threatened. Instead of open benchmarks and shared weights, we now see coordinated rate‑limiting, anomaly detection on API usage, and potentially legal pressure on anyone suspected of large‑scale distillation. The fact that these efforts are framed largely in terms of keeping up with China gives them political cover, even when they’re just as useful for keeping out domestic upstarts.
3. Consolidation at the top of the stack
Third, Musk’s own ranking of Anthropic, OpenAI and Google as the clear leaders underscores how concentrated the frontier has become. If those three (plus a handful of Chinese players) also succeed in making distillation hard or legally risky, the barrier to entry for new general‑purpose model labs will rise even further. That would push startups toward specialised, smaller models and creative fine‑tuning rather than trying to build foundation models from scratch.
Historically, we’ve seen something similar in cloud computing: once AWS, Azure and Google Cloud reached critical scale, new entrants stopped trying to match them on raw infrastructure and instead built on top. Distillation defences may cement a comparable "three‑and‑a‑half‑player" structure at the foundation‑model level.
5. The European / regional angle
For Europe, Musk’s admission is less about celebrity drama and more about regulatory timing.
The EU AI Act, which will start applying gradually in the middle of this decade, puts strong emphasis on training data transparency, model documentation and systemic risk for "general‑purpose AI". Distillation doesn’t fit neatly into those categories, but regulators in Brussels will be watching closely. If a European lab distils capabilities from a US model whose terms of service forbid it, is that a contract issue, an IP issue, or a matter for competition law? Expect test cases before 2030.
European developers are in a bind. On one hand, distillation is one of the few realistic techniques that could let smaller players compress expensive frontier models into cheaper, locally‑hosted systems that respect EU data‑protection expectations. On the other, any explicit reliance on distillation could conflict with ToS and complicate compliance with the AI Act’s transparency and governance rules.
There’s also a sovereignty angle. If US and Chinese giants both clamp down on distillation, European industry may be forced into long‑term dependency on foreign APIs. That sits uneasily alongside the EU’s digital‑sovereignty ambitions and the Commission’s broader push, via the Digital Markets Act and Digital Services Act, to reduce structural dependence on a few US platforms.
Europe still has a chance to carve out a distinctive position: clarifying what types of model‑on‑model training are acceptable, under what safeguards, and perhaps even treating some forms of distillation as pro‑competitive rather than parasitic.
6. Looking ahead
Several things are likely over the next two to three years.
1. Technical "model DRM" arms race
Expect a quiet boom in techniques that try to detect and frustrate distillation: output watermarking, unusual‑query detection, and pricing models that make bulk querying uneconomical. As with DRM in music and video, these defences will be leaky, but they will raise the cost and legal risk of straightforward scraping.
2. Legal test cases and ToS hardening
Musk’s testimony almost invites OpenAI and others to tighten their contracts and pursue at least one high‑profile case to set an example. Even if the first lawsuits settle, they will put lawyers and regulators on notice that model‑on‑model training is no longer a purely academic topic.
3. More closed, vertically integrated AI stacks
If distillation becomes harder, the rational move for smaller companies may be to align with one ecosystem (OpenAI/Microsoft, Anthropic/Amazon, Google, or a Chinese stack) rather than betting on independent foundation‑model development. In Europe, that means more pressure on policymakers to support indigenous capabilities through public compute, research funding and perhaps a more permissive stance on certain forms of distillation.
4. Reputational risk and hypocrisy narratives
Finally, Musk has handed critics of all major labs an easy talking point: nearly everyone in this space has benefited from loose attitudes to others’ data or models at some point. The more loudly incumbents condemn distillation from China or startups, the more they will be asked what they did themselves on the way up.
7. The bottom line
Musk’s acknowledgment that xAI trained Grok in part on OpenAI models doesn’t just embarrass one company; it exposes a structural tension in the AI economy. The same labs that grew by aggressively harvesting external data now want strong property rights over the behaviours of their own models. Whether regulators treat distillation as theft, fair competition or something in between will shape who can challenge today’s AI giants. The real question for readers and policymakers alike: do we want AI capabilities to behave more like public knowledge or like tightly controlled corporate IP?
