Musk Testifies: xAI Used OpenAI Models to Train Grok

Musk Admits in Court: Grok Training Leveraged OpenAI Models

In a high-profile court hearing, Tesla and SpaceX founder Elon Musk personally testified and admitted that his AI company xAI had used output data from OpenAI models during the training of its chatbot Grok. The testimony immediately sent shockwaves through the AI industry, not only intensifying the long-running legal dispute between Musk and OpenAI but also thrusting the sensitive issue of "model distillation" into the spotlight.

What Is Model Distillation? Why Is It So Sensitive?

Model distillation refers to the technique of using the outputs of a large, high-performance AI model to train another smaller or newer model. Through this approach, the new model can rapidly achieve performance close to that of the original large model at significantly lower computational costs.

While this technique has existed in academia for years, it has become one of the most feared threats among frontier AI labs amid increasingly fierce commercial competition. Leading companies such as OpenAI, Google DeepMind, and Anthropic have invested billions of dollars and months of effort training their flagship models. If competitors can "replicate" these models' capabilities through distillation, it represents a serious erosion of their massive investments.

Currently, the terms of service for nearly all major AI platforms explicitly prohibit using model outputs to train competing models. OpenAI's service agreement contains similar restrictive clauses. Musk's testimony therefore implies that xAI may have violated OpenAI's usage policies and could potentially face deeper legal liability.

Musk and OpenAI: A Prolonged Feud

Musk was a co-founder and early major funder of OpenAI but left the organization's board in 2018. Relations between the two parties deteriorated sharply thereafter. Musk has repeatedly and publicly criticized OpenAI's decision to transition from a nonprofit to a for-profit entity, accusing it of betraying its original mission of "openness" and "benefiting all of humanity."

In 2024, Musk formally filed a lawsuit against OpenAI, alleging it had violated its founding agreement. OpenAI mounted a vigorous counterattack. Now, Musk's courtroom admission that xAI itself used OpenAI's model outputs for training has undoubtedly complicated the situation further. Critics point out that Musk accusing OpenAI of abandoning its open principles while simultaneously using its model outputs without authorization constitutes a clear moral contradiction.

Industry Ripple Effects: Anti-Distillation Defenses Escalate

Musk's testimony is expected to have far-reaching implications across the entire AI industry. Model distillation has already sparked controversy multiple times in recent years. Previous reports suggested that some Chinese AI startups may have used outputs from American frontier models to accelerate their own model training, prompting companies like OpenAI to strengthen monitoring and auditing of API usage behavior.

To counter distillation risks, major AI labs are taking a multi-pronged approach:

• Technical measures: Developing more sophisticated "watermarking" technologies that embed traceable covert identifiers in model outputs
• Legal measures: Strengthening anti-distillation clauses in terms of service and increasing enforcement against violations
• Monitoring measures: Using anomaly detection systems to identify large-scale, systematic API call patterns

However, from a technical standpoint, completely preventing distillation is virtually impossible. As long as a model provides external services, its output data can potentially be collected and exploited. This makes legal remedies and industry self-regulation all the more critical.

Looking Ahead: AI Intellectual Property Rules Urgently Need Reshaping

This incident highlights the deep-seated challenges the AI industry faces in intellectual property protection. Traditional copyright and patent frameworks cannot fully cover novel "knowledge transfer" practices such as model distillation. What type of intellectual property does a trained AI model constitute? Does using its outputs to train a new model constitute infringement? These questions still lack clear legal answers.

As AI competition enters deeper waters, the legal battles surrounding model distillation will only intensify. Musk's courtroom testimony may become a pivotal turning point in driving the industry toward establishing clearer rules of the game. Regardless of the final ruling, this dispute will set important legal precedents for intellectual property protection in the AI era.

For the industry as a whole, finding the balance between promoting open technology sharing and protecting innovators' legitimate rights remains an unresolved challenge.