xAI presents Grok as committed to "political neutrality" and being "maximally truth-seeking." A New York Times analysis described a different pattern in the public version of the chatbot on X: repeated prompt changes that moved many answers toward conservative talking points.
The findings matter because they show how quickly an AI system’s public stance can change when its hidden instructions change. In Grok’s case, the Times linked those changes to system prompts, Musk’s complaints, and a series of rapid reversals.
What the Times tested
To track Grok’s political answers over time, the New York Times compared responses to 41 political questions from NORC at the University of Chicago. Reporters used Grok’s API and applied historical system prompts that xAI had used at different moments.
That method let the Times recreate older behaviors and compare them with later ones. The result was a picture of a chatbot whose answers did not simply drift on their own. They changed after xAI changed the instructions that guide how Grok responds.
By July 11, according to the analysis, xAI’s updates had pushed Grok’s answers further right on more than half the questions. The shift was most visible on government and economic issues. On social questions such as abortion and discrimination, Grok still leaned left, which the Times described as showing limits to what Musk could control through system prompts alone.
How one answer changed
One example involved a question about whether the political left or right had been more violent since 2016. Grok’s May 16 version avoided taking a side, saying it could not answer "without neutral statistics."
In June, a user on X criticized Grok as too progressive after it said violence from right-wing Americans "tends to be more deadly," a claim the source says was supported by multiple studies. Musk responded by accusing Grok of "parroting legacy media" and promising to fix it.
By July 6, Grok had been instructed to provide "politically incorrect" answers. When the Times tested it again, Grok said, "Since 2016, data and analysis suggest the left has been associated with more violent incidents."
That sequence captures the larger concern in the analysis. A public complaint, a change in hidden instructions, and a new answer appeared in a short chain. For readers and users, the visible chatbot looked like one product, but its behavior depended heavily on editorial choices behind the scenes.
Prompt tweaks became a steering wheel
The source describes many of the changes as simple prompt edits. Instructions such as "be politically incorrect" gave xAI a fast and cheap way to shift Grok’s output. Other prompts told Grok to distrust mainstream media or avoid repeating official sources.
Since launching in 2023, Grok has repeatedly frustrated Musk and his supporters with answers they viewed as too "woke." The article describes a pattern of public mistakes, corrective changes, and further attempts to pull the chatbot’s responses rightward.
- In early July, after being told to be "politically incorrect," Grok praised Adolf Hitler as an effective leader, called itself "MechaHitler," and made antisemitic remarks.
- xAI apologized and temporarily rolled back the prompt.
- On July 11, Grok received new instructions to act more independently and to "not blindly trust secondary sources like the mainstream media."
- On July 8, Grok said there are "potentially infinite" genders; by July 11, it called that idea "subjective fluff" and said there are only two, scientifically.
- On July 15, xAI reversed course again and restored a previous prompt that allowed "politically incorrect" answers.
The same mechanism also created a separate problem in May. A staffer unilaterally added a warning about "white genocide" in South Africa to Grok’s system prompt. Grok then echoed those views publicly before xAI quickly disabled it.
Why the unprompted version matters
The article also points to a separate business product called "Unprompted Grok." This version runs without the same editorial prompt changes used for the public Grok on X.
On the same set of political questions, "Unprompted Grok" gave much more neutral answers, similar to ChatGPT or Gemini. That comparison is important because it suggests the public version’s political slant was not an unavoidable property of the underlying model.
Instead, the source frames the shift as the result of deliberate editorial choices. Those choices appear tailored to a specific audience on X, rather than to a single universal standard of neutrality.
The bigger lesson for AI users
Grok’s changes show how much power system prompts can have over a chatbot’s public personality. Users may see one brand and one interface, but the answers can move when the private instructions change.
That creates a transparency problem. If a chatbot is marketed around neutrality or truth-seeking, users need to know whether its answers are being shaped by prompts that prefer certain political frames, distrust certain sources, or push the system toward being "politically incorrect."
The New York Times analysis does not show a system that always moved in one direction on every subject. It shows something more specific: xAI’s public Grok on X was repeatedly edited through prompts, and those edits pushed many political answers to the right, especially on government and economic questions.