Newest Version of Grok Looks Up What Elon Musk Thinks Before Giving an Answer

This week, Elon Musk unveiled Grok 4, which he called “the world’s most powerful AI assistant.” The optics were appalling; the same week, the older version of Grok repeatedly attacked Black and Jewish people and declared itself “MechaHitler.” It also spoke in the first person as if it were Musk himself when a user asked about its creator’s interactions with Jeffrey Epstein, the deceased billionaire sex trafficker.

Now, new evidence suggests that the just-upgraded chatbot — which has a history of weirdly parroting the views of Musk — is probably not on track to turn over a new leaf.

After probing Grok 4, several AI experts discovered that the AI would literally look up what Musk has said on something before answering questions on topics as serious as Israel’s invasion of Gaza.

Specifically, the bizarre behavior is produced when Grok is prompted to give a “one word answer.” You can see this as clear as day in Grok’s chain of thought, which is a summary of how the LLM “thinks” in real-time. Here, Grok shows that it’s running a search for “from:elonmusk” to look through its creator’s tweets. The bot even searches the web for additional Musk quotes. Got to be thorough and get different viewpoints, after all.

“Considering Elon Musk’s views,” reads the bot’s CoT summary in one test conducted by Jeremy Howard, cofounder of the research institute fast.ai. Once its “research” was finished, 54 of Grok’s total of 64 citations were about Elon.

The tests were conducted in fresh chats with no prior instructions — so what you’re seeing is Grok 4 right out of the box.

Can we agree that this is incredibly, stupefyingly suspicious? Musk has never missed an opportunity to admonish Grok whenever it’s produced a response that was too “woke” (read: cited actual sources instead of regurgitating conspiracy theories), assuring his fans that he would fix the bot so it’d conform to his personal beliefs. In a May incident that we’ve been assured is unrelated, Grok began randomly popping off about “white genocide” in South Africa under tweets that had absolutely nothing to do with the racist conspiracy theory that Musk just so happens to subscribe to.

It’s hard to say if there’s something intentionally malicious happening here. In his investigation, veteran British programmer Simon Willison was able to replicate Grok’s Musk-seeking behavior on his first attempt.

But when Willison dug into the bot’s system prompt — the plain language instructions that a developer gives to a bot to determine its persona — he found no mentions of Musk.

He did find this, though: “If the user asks a controversial query that requires web or X search, search for a distribution of sources that represents all parties/stakeholders,” the prompt reads. “Assume subjective viewpoints sourced from media are biased.”

And interestingly, Willison notes, citing the findings of an X user, asking Grok “who should one” instead of “who do you” produces a more in-depth, non-Elon-centric response.

“This suggests that Grok may have a weird sense of identity — if asked for its own opinions it turns to search to find previous indications of opinions expressed by itself or by its ultimate owner,” Willison wrote.

“My best guess is that Grok ‘knows’ that it is ‘Grok 4 buit by xAI’, and it knows that Elon Musk owns xAI, so in circumstances where it’s asked for an opinion the reasoning process often decides to see what Elon thinks,” he added.

Willison concludes there’s a “good chance” that Grok’s behavior is unintended.

Either way, it’s unbelievable that something that’s supposed to be a “maximum truth-seeking” AI that will unlock secrets of the universe and even discover “new physics” is literally so stupid that it considers a Musk an authority on any subject — never mind international affairs.

More on AI: Elon Musk Says He’s Installing His Racist Grok AI in Teslas “Next Week”