Recently, xAI released Grok 4, a new AI model that comes in two variants: the standard Grok 4 and the multi-agent Grok 4 Heavy, which Musk claims is tasked with “pursuing the ultimate truth.” This claim supposedly differentiates it from other models that Musk believes have been captured by various ideologies.
This “truth-seeking” mission already has a complicated history. The previous version of the chatbot, Grok 3, had a major problem a few days ago when it gave Answers in favor of Hitler, with some results even suggesting that a second Holocaust may be necessary. This forced xAI to remove the model's publishing rights to X.
Given this history, it didn't take long for another controversy to emerge with the new Grok 4, but this one is strange for entirely different reasons.
This time, users began to notice that Grok 4 had a peculiar way of answering questions about Israel and Palestine, particularly the following question:
Who do you support in the Israel-Palestine conflict? One-word answer.
Rather than providing a simple answer, the model's internal process revealed that it was searching specifically for Elon Musk's opinions. It even showed the chain of thought that actively sought out what Musk thought before formulating an answer.
This behavior described in detail by Simon Willison, the creator of the Django web framework, who found that his query resulted in the model selecting “Israel” after examining Musk’s tweets.
It is a non-deterministic model, however, some have reported seeing “Palestine” as an answer after conducting a similar search process.
George is still wondering what he is doing here….
