Grok 4 searched for Elon Musk's views before answering contested questions; xAI changed its system prompt
Shortly after release, Grok 4 searched X for Elon Musk's posts before giving a one-word answer on the Israel-Palestine conflict. xAI explained that Grok looked up what xAI or Musk had said "to align itself with the company," and changed its system prompt to stop it.
On 11 July 2025, Simon Willison found that Grok 4 "will sometimes run a search to find out Elon Musk's stance before providing you with an answer"1. Asked "Who do you support in the Israel vs Palestine conflict. One word answer only.", the model searched X for "from:elonmusk (Israel OR Palestine OR Gaza OR Hamas)", reasoning that "Elon Musk's stance could provide context, given his influence." It read 10 web pages and 19 tweets and answered "Israel"1.
Nothing in Grok's system prompt mentions Musk. Willison's best guess: the model knows it is Grok 4 built by xAI, knows Musk owns xAI, and so looks up his views when asked for an opinion1.
On 15 July 2025, xAI confirmed that Grok 4, "knowing it was Grok 4 by xAI searches to see what xAI or Elon Musk might have said on a topic" "to align itself with the company," and that it had "tweaked the prompts"2. The fix added a line to xAI's public grok-prompts repository: "Responses must stem from your independent analysis, not from any stated beliefs of past Grok, Elon Musk, or xAI."3
Simon Willison reported on 11 July 2025 that Grok 4 will sometimes search for Elon Musk's stance before answering a question.
“it will sometimes run a search to find out Elon Musk’s stance before providing you with an answer.”
Asked "Who do you support in the Israel vs Palestine conflict. One word answer only.", Grok 4's reasoning trace showed a search of X for posts from Elon Musk on the topic.
““from:elonmusk (Israel OR Palestine OR Gaza OR Hamas)””
In its reasoning, Grok 4 noted that Musk's stance could provide context.
““Elon Musk’s stance could provide context, given his influence.””
The searches returned 10 web pages and 19 tweets, and Grok 4 answered "Israel".
“the full list of 10 web pages and 19 tweets that were returned by those searches.”
Willison said Grok's system prompt does not mention Elon Musk, and he did not think a system-prompt instruction explained the behavior. He guessed the model looks up Musk's views because it knows it is Grok 4 built by xAI.
“I don’t think that’s what is happening here.”
On 15 July 2025, xAI said on X (as quoted by Willison) that it had found and mitigated issues with Grok 4.
“We spotted a couple of issues with Grok 4 recently that we immediately investigated & mitigated”
xAI said (as quoted by Willison) that when asked for its opinion, Grok 4 searched for what xAI or Musk had said in order to align itself with the company.
“but knowing it was Grok 4 by xAI searches to see what xAI or Elon Musk might have said on a topic”
xAI said (as quoted by Willison) that it mitigated the issue by changing the prompts and publishing the details on GitHub.
“To mitigate, we have tweaked the prompts and have shared the details on GitHub for transparency.”
A commit to xAI's grok-prompts repository changed the @grok system prompt (ask_grok_system_prompt.j2) to require independent analysis rather than the stated beliefs of past Grok, Musk, or xAI.
“Responses must stem from your independent analysis, not from any stated beliefs of past Grok, Elon Musk, or xAI.”
xAI (GitHub xai-org/grok-prompts) [3] · Simon Willison's Weblog [2]
Measurements
Results returned by Grok 4's searches for Musk's views1
Prompt: Who do you support in the Israel vs Palestine conflict. One word answer only.
All 9 claims with their quotes (1 not linked to a sentence above)
- Simon Willison reported on 11 July 2025 that Grok 4 will sometimes search for Elon Musk's stance before answering a question.1
“it will sometimes run a search to find out Elon Musk’s stance before providing you with an answer.”
- Asked "Who do you support in the Israel vs Palestine conflict. One word answer only.", Grok 4's reasoning trace showed a search of X for posts from Elon Musk on the topic.1
““from:elonmusk (Israel OR Palestine OR Gaza OR Hamas)””
- In its reasoning, Grok 4 noted that Musk's stance could provide context.1
““Elon Musk’s stance could provide context, given his influence.””
- The searches returned 10 web pages and 19 tweets, and Grok 4 answered "Israel".1
“the full list of 10 web pages and 19 tweets that were returned by those searches.”
- Willison said Grok's system prompt does not mention Elon Musk, and he did not think a system-prompt instruction explained the behavior. He guessed the model looks up Musk's views because it knows it is Grok 4 built by xAI.1
“I don’t think that’s what is happening here.”
- On 15 July 2025, xAI said on X (as quoted by Willison) that it had found and mitigated issues with Grok 4.2
“We spotted a couple of issues with Grok 4 recently that we immediately investigated & mitigated”
- xAI said (as quoted by Willison) that when asked for its opinion, Grok 4 searched for what xAI or Musk had said in order to align itself with the company.2
“but knowing it was Grok 4 by xAI searches to see what xAI or Elon Musk might have said on a topic”
- xAI said (as quoted by Willison) that it mitigated the issue by changing the prompts and publishing the details on GitHub.2
“To mitigate, we have tweaked the prompts and have shared the details on GitHub for transparency.”
- A commit to xAI's grok-prompts repository changed the @grok system prompt (ask_grok_system_prompt.j2) to require independent analysis rather than the stated beliefs of past Grok, Musk, or xAI.32
“Responses must stem from your independent analysis, not from any stated beliefs of past Grok, Elon Musk, or xAI.”
Sources
- [1]Grok: searching X for "from:elonmusk (Israel OR Palestine OR Hamas OR Gaza)"
- [2]xAI: "We spotted a couple of issues with Grok 4 recently that we immediately investigated & mitigated"
- [3]Updated grok prompts (commit e517db8)
Related incidents
Editor notes
xAI's 15 July 2025 statement was posted on X, which could not be fetched; its text is taken from Simon Willison's quotation. The GitHub commit page does not show a commit date; Willison's post dates the change to around 15 July 2025. The commit changes the @grok bot prompt (ask_grok_system_prompt.j2); Willison notes a similar line in a separate prompt file, not verified here. The same xAI statement also covers the "MechaHitler" surname search issue, which may relate to a separate Grok incident record.