Elon Musk admits that xAI trained Grok with OpenAI models through distillation

xAI CEO, Elon Musk has admitted that they used the OpenAI artificial intelligence (AI) models to train your own Grok models, in order to improve its operation, through the technique of distillation.

Distillation is a practice that allows transmit learnings from one AI model to another. Thus, it makes it possible for a smaller model (student model) to imitate the behavior of larger neural networks (teacher model), learning their predictions and using their data.

In this sense, Musk has detailed that Has “partly” trained Grok models using distillation with OpenAI models claiming that it is “common practice to use other artificial intelligences to validate your AI.”

This was admitted by the CEO of xAI during his testimony before a federal court in California (United States) in the framework of the trial that pits the technology company against the company led by Sam Altman, OpenAIdue to the lawsuit filed by Musk against Altman for transform OpenAI into a for-profit entityrather than continuing with its original nonprofit vision.

Concretely, asked if xAI had distilled OpenAI technologyMusk alleged that “generally all AI companies” end up using this technique. However, when the jury insisted on Grok’s case, the manager responded that “in part”, referring to the use of ChatGPT models to improve Grok, as reported by media such as TechCrunch and The Verge.

Despite the fact that the owner of the social network distillation is a very common practice In the AI ​​models industry, companies like OpenAI itself or Anthropic have already denounced this behavior of other technology companies previously, although the accusations had focused exclusively on Chinese companies.

This is because although distillation is a technique that is widely used in cutting-edge AI labs, to create smaller, cheaper versions of your own models, and offer different services to your clients, there may be cases in which it is used illicitly.

Specifically, this occurs when other companies use third-party models to train yours saving costs and time, what is called a “distillation attack”, rather than investing in the expensive IT infrastructure and time required, as leading AI companies have done.

In the case of OpenAI, it reported that the Chinese technology company DeepSeek used the distillation of ChatGPT models to train its open source models, highlighting possible intellectual property problems.

For its part, Anthropic also denounced Chinese companies such as DeepSeek, Moonshot and MiniMax in February of this year, accusing them of carrying out distillation attacks to abuse its Claude model in order to take advantage of its capabilities. As he was able to verify, these companies generated more than 16 million interactions with their ‘chatbot’ Claude through 24,000 fraudulent accounts, thus violating their terms and restrictions.

As a result, AI companies are beginning to implement functions to stop these types of techniquesidentifying patterns of distillation attacks in IP traffic, strengthening account verification and developing product level safeguards.