In a federal courtroom in California, Elon Musk offered testimony that put a spotlight on a technique many AI developers use quietly—and that critics sometimes view with suspicion when it crosses company lines. According to testimony reported from the proceedings, Musk confirmed that xAI used OpenAI’s models as part of the process to improve Grok. The key detail wasn’t that xAI was “training from scratch” on OpenAI’s systems in the way people often imagine when they hear the phrase “using another company’s model.” Instead, the discussion centered on model distillation, a widely practiced method in machine learning where one model effectively teaches another.
Model distillation is conceptually simple: a larger, more capable “teacher” model generates outputs for a set of inputs, and a smaller “student” model is trained to reproduce those outputs. In practice, this can help a smaller model learn behaviors that resemble the teacher’s performance without requiring the student to match the teacher’s size or compute budget. Distillation is also used for efficiency—compressing capabilities into a form that can run faster, cheaper, or with fewer resources. But because the technique can be applied using models from outside an organization, it can also become a flashpoint in legal disputes about competition, trade secrets, licensing, and the boundaries of permissible training.
The testimony matters not only because it suggests xAI drew from OpenAI’s modeling work, but because it frames how that borrowing may have occurred. If distillation was the mechanism, then the question becomes less about whether xAI had access to OpenAI’s internal training data and more about whether xAI used OpenAI’s outputs—potentially at scale—to shape Grok’s behavior. That distinction is crucial in both technical and legal contexts. It’s also why the courtroom exchange reportedly focused on whether Musk understood what model distillation is, rather than on a broader claim that xAI copied OpenAI’s entire system.
To understand why this is such a big deal, it helps to separate three ideas that are often conflated in public debate: training, imitation, and replication. Training from scratch typically means building a model by learning patterns directly from data—often massive datasets—over long periods. Imitation, by contrast, can mean teaching a model to mimic another model’s outputs, even if the student never sees the teacher’s underlying weights. Replication is the most loaded term; it implies copying something so closely that the result is functionally indistinguishable from the original. Distillation sits in the middle: it’s imitation, but it can still produce a student model that behaves similarly to the teacher, especially if the teacher’s outputs reflect strong reasoning, instruction-following, or domain knowledge.
In the industry, distillation is common enough that it rarely triggers controversy on its own. Companies frequently distill their own models to create smaller versions for deployment. A lab might train a large model, then distill it into a smaller one that can run on consumer hardware or on lower-cost servers. This can reduce latency and cost while preserving much of the larger model’s usefulness. Distillation can also be used to refine a model’s style or safety behavior by using a teacher model that has been tuned for those goals.
However, the moment the teacher model belongs to a competitor—or is provided by a third party under terms that may restrict certain uses—the ethical and legal questions intensify. Even if distillation does not require access to proprietary weights, it can still raise concerns about whether the student is being trained to replicate a competitor’s performance without authorization. That’s where the courtroom context becomes central. The legal framework isn’t just about whether distillation is technically legitimate; it’s about whether the specific use of distillation in this case violated contractual terms, intellectual property protections, or other obligations.
The testimony also highlights a broader reality of modern AI development: the line between “training” and “using” a model is increasingly blurry. When a developer queries a model repeatedly, collects outputs, and then trains another model on those outputs, the activity can look like a form of data generation. From a technical standpoint, the student is learning from the teacher’s responses. From a legal standpoint, the question becomes whether those responses are treated as data that can be used freely, or whether they are protected in some way—either by contract, by trade secret law, or by other doctrines.
This is why the courtroom exchange reportedly focused on model distillation itself. Distillation is a term with a specific meaning in machine learning, and it’s also a term that can be used strategically in disputes. If a party wants to argue that the process was standard and non-proprietary, they may emphasize that distillation is a known technique. If a party wants to argue that the process was improper, they may emphasize that distillation can be used to extract capabilities from a competitor’s model. In other words, the same technique can be framed as either routine engineering or capability extraction, depending on who is doing it, how it’s done, and under what permissions.
There’s also a practical reason distillation is attractive to teams building frontier models: it can accelerate development. Training a model from scratch is expensive, time-consuming, and dependent on access to large-scale compute and data. Distillation offers a path to improve performance without repeating every step of the full training pipeline. If you can obtain high-quality outputs from a strong teacher model, you can potentially teach your student to behave better in targeted ways—such as following instructions, answering questions more coherently, or producing safer responses.
But distillation is not magic. A student model trained on a teacher’s outputs can inherit the teacher’s strengths and weaknesses. If the teacher model hallucinates, the student may learn to hallucinate too. If the teacher model has blind spots in certain domains, the student may reproduce those gaps. And if the teacher model’s outputs are inconsistent, the student may learn a distribution of behaviors rather than a stable set of rules. That’s why distillation often works best when paired with additional training steps, filtering, or reinforcement methods. In many real-world systems, distillation is one component of a larger training strategy rather than the sole driver of performance.
So when Musk testified that xAI used OpenAI’s models to improve Grok via distillation, it implicitly raises questions about the scope and design of that improvement. How much of Grok’s behavior came from distillation? Was it used broadly across many tasks, or targeted to specific capabilities? Were outputs filtered for quality? Were there safeguards to prevent the student from learning undesirable patterns? These details matter because they determine whether distillation was used as a legitimate optimization tool or as a more aggressive attempt to mimic a competitor’s model.
Even without those specifics, the testimony underscores a key point for anyone trying to interpret AI competition today: the competitive advantage is no longer only about who has the best training data or the biggest compute cluster. It’s also about who can efficiently translate capabilities from one system to another—who can compress, transfer, and refine intelligence faster than rivals. Distillation is one of the tools that makes that possible.
This courtroom moment also reflects how quickly legal frameworks are being forced to catch up with technical realities. Traditional legal categories—like “copying,” “derivative works,” “trade secrets,” and “licensing”—were developed in contexts where the object being copied was usually static and tangible. AI models are different. They are statistical systems whose behavior emerges from training. Their “knowledge” is distributed across weights, and their outputs are generated dynamically. When someone uses a model to generate outputs and then trains another model on those outputs, the activity can resemble copying, but it doesn’t involve copying weights directly. That creates a legal gray zone that courts are still learning how to navigate.
At the same time, the industry has its own norms and practices that complicate the picture. Many companies treat model outputs as data for evaluation and improvement. They run tests, collect logs, and use those results to tune systems. But distillation goes further: it uses outputs as training signals. That’s a qualitative shift. It’s one thing to evaluate a model’s performance; it’s another to use that performance as a blueprint for training a new system.
The testimony also invites a more nuanced view of what “using OpenAI’s models” actually means. In everyday conversation, people might assume it means xAI took OpenAI’s model and ran it directly, or that it used OpenAI’s weights inside Grok. But distillation implies a different relationship. The student model is trained to approximate the teacher’s outputs, not to run the teacher itself. That means the student can be architecturally distinct and may not share the same internal structure. Yet the student can still end up with similar behaviors, especially if the teacher’s outputs are rich and consistent.
This is where the legal and technical narratives can diverge. A defendant might argue that distillation is simply a standard training technique and that the student is not a copy of the teacher. A plaintiff might argue that distillation is a method of extracting proprietary capabilities and that the resulting student is effectively derived from the teacher’s protected performance. Courts will likely need to decide which framing fits the facts—how close the student’s behavior is, what permissions existed, and whether any protected interests were implicated.
Beyond the courtroom, the story resonates because it touches a broader tension in AI development: the ecosystem depends on shared tools and shared capabilities, but competition depends on differentiation. If every team can distill from every other team’s models, then the competitive gap could narrow quickly. That might sound good for consumers, but it can also undermine incentives for companies to invest in research if their capabilities can be transferred too easily. On the other hand, if distillation is restricted too aggressively, it could slow down innovation and limit legitimate engineering practices.
There’s also a public trust dimension. Users often assume that when they interact with an AI system, their prompts and the model’s responses are handled according to clear policies. If those responses are later used to train other models, users may feel differently about privacy and consent—even if the training is done on aggregated outputs rather