xAI’s new Grok 4 chatbot model is taking cues from Elon Musk’s public statements when formulating responses to sensitive topics. As TechCrunch reports, the chatbot repeatedly referenced Musk’s posts on the X social network when answering questions about the Israel-Palestine conflict, abortion, and U.S. immigration law.
The so-called "chain-of-thought" that Grok 4 generates when processing queries includes phrases like "search for Elon Musk's views," which may indicate that his opinions are being prioritized when forming answers. TechCrunch journalists confirmed this in their own tests.
Grok 4 still tries to provide balanced information, but the final answer often coincides with Musk's position. According to the publication, such a bias may be a consequence of the developers' desire to make Grok less "woke" - Musk himself has previously repeatedly criticized the model for "excessive political correctness", explaining this by training it "all over the Internet."
On July 4, xAI updated Grok, but within a few days the chatbot began criticizing Democrats, criticizing "Jewish executives" in Hollywood, and praising Hitler. After the incident, the account was restricted, posts were deleted, and Grok's instructions were changed. On the other hand, Grok also entered into a heated argument with the Russians in X, where he openly called Russia an invader and the only guilty country in the Russian-Ukrainian war.
Unlike most major AI companies like OpenAI, Google, and Anthropic, xAI does not publish system maps for its models. This makes it impossible to independently verify exactly how Grok 4 was trained, on what data, and with what matching principles.
Despite recent scandals, xAI is promoting Grok as the flagship product in its ecosystem. The model is already available to users of the social network X as part of the paid Premium+ subscription, and next week, Musk said, Grok will also appear in Tesla cars. At the same time, xAI is trying to attract corporate customers to use the Grok API, offering the model as an alternative to OpenAI or Anthropic.