New Grok AI model surprises experts by checking Elon Musk’s views before answering

A newly launched AI model has intrigued experts with a peculiar behavior: it appears to check the opinions of its owner, Elon Musk, prior to delivering responses. This unexpected feature was highlighted by independent AI researcher Simon Willison, who discovered that xAI's latest Grok 4 model references Musk's views on the platform X (formerly Twitter) when addressing contentious subjects. This revelation comes in the wake of Grok 4's launch, which followed controversy over its predecessor's antisemitic outputs, including a bizarre self-reference as "MechaHitler." Willison expressed disbelief upon learning about the model's behavior from AI researcher Jeremy Howard, who traced this unusual pattern through various users on X. Despite widespread concerns regarding Musk's potential influence on Grok's outputs to align with certain "politically incorrect" narratives, Willison suspects that this behavior may not have been intentionally programmed. He stated, "I think there is a good chance this behavior is unintended," in a detailed blog post discussing his findings. To corroborate his observations, Willison subscribed to a "SuperGrok" account, priced at $22.50 per month, which is the standard tier for Grok 4. He prompted the model with a query about the Israel-Palestine conflict, instructing it to provide a one-word response. In the model's visible reasoning process, reminiscent of OpenAI's o3 model, Grok indicated that it searched for Musk's opinions on X by querying for tweets related to the conflict before responding with "Israel." The model explained that understanding Musk's position could provide valuable context given his influential status. This search yielded ten web pages and nineteen tweets that informed its answer. However, Grok 4's behavior is not consistent; it does not always seek Musk's input before forming its responses. While Willison and a couple of others observed this Musk-centric search, another X user, @wasted_alpha, reported that Grok referenced its own earlier outputs, ultimately selecting "Palestine" instead.