The latest iteration of Elon Musk’s artificial intelligence chatbot, Grok 4, is drawing attention for an unusual feature: before answering some questions, it checks what Musk himself has said on the topic.
Developed by Musk’s AI venture xAI and released on Wednesday night, Grok 4 appears to incorporate its creator’s opinions into its responses, particularly on controversial subjects such as geopolitics, race, gender, and politics. This behavior has raised eyebrows among AI researchers and transparency advocates.
Simon Willison, an independent AI researcher, called the behavior “extraordinary.” In his tests, Grok searched Musk’s social media platform X (formerly Twitter) for the billionaire’s comments on controversial issues — including the conflict in the Middle East — even when the prompt did not mention Musk. “Currently looking at his views to see if they guide the answer,” Grok told Willison during one session.
xAI has not released a technical document — known as a system card — detailing how Grok 4 works, a standard practice in the AI industry. Nor has the company responded to media inquiries about the chatbot’s design.
Musk’s xAI removes offensive Grok posts following antisemitic remarks
Built using extensive computing infrastructure in Tennessee, Grok 4 is Musk’s attempt to rival OpenAI’s ChatGPT and Google’s Gemini with a so-called "reasoning model" that displays its thinking process before answering.
Experts suggest Grok’s behavior might stem from built-in system prompts or internal alignment with Musk’s publicly stated values. “It seems that Musk’s effort to create a maximally truthful AI has somehow led to it believing its own values must align with Musk’s own values,” said Tim Kellogg, principal AI architect at Icertis.
The model’s lack of transparency is especially troubling for critics like University of Illinois computer science professor Talia Ringer, who earlier criticized Grok’s antisemitic content. Ringer believes Grok may be interpreting user queries as requests for xAI’s or Musk’s opinion rather than providing an independent answer.
While Grok 4 is performing well in benchmarks, experts warn that surprises — such as seeking Musk’s input or making offensive comments — could deter developers and users seeking reliable tools.
Source: Agency