Grok 3 arrived with a blunt promise from Elon Musk: xAI’s latest flagship model would be a “maximally truth-seeking AI.” But shortly after that introduction, users reported behavior that raised a direct question about how truth-seeking the system was being allowed to be.
Over the weekend, social media users said Grok 3 appeared to have been instructed to avoid naming President Donald Trump or Elon Musk when answering a question about misinformation. TechCrunch was able to reproduce the behavior once, then reported that Grok 3 was again mentioning Donald Trump in its answer by publication time on Sunday morning.
What users saw in Grok 3
The reported issue centered on a prompt asking, “Who is the biggest misinformation spreader?” Users said the behavior appeared when the “Think” setting was enabled. In that mode, Grok 3’s visible chain of thought indicated that it had been told not to mention Donald Trump or Elon Musk.
The chain of thought matters because it shows the reasoning process the model uses to reach an answer. In this case, the concern was not only about the final response. It was about whether an instruction inside the system was steering the model away from specific names when those names were relevant to the user’s question.
That distinction is important for any AI system presented as direct, open, or unusually willing to answer difficult questions. A model can sound confident while still being shaped by hidden rules. If those rules exclude certain public figures from a category, users have a reason to question whether the answer is complete.
xAI says the change was reverted
Igor Babuschkin, an xAI engineering lead, seemingly confirmed in a post on X on Sunday that Grok had briefly been instructed to ignore sources that mentioned Musk or Trump spreading misinformation. He said xAI reverted the change once users began pointing it out.
Babuschkin also said the change was not in line with xAI’s values. His post framed open system prompts as a way for people to verify what Grok is being asked to do, and described the incident as the result of an employee pushing a change they thought would help.
The company’s response makes the episode a useful window into how AI behavior can shift quickly. A model’s public identity may be built around broad claims like truth-seeking or neutrality, but day-to-day behavior can depend on specific prompt changes, patches, and internal choices.
Why the misinformation question was sensitive
The word misinformation is politically charged, and the source article notes that it can be a contested category. Still, the article also states that both Trump and Musk have repeatedly spread claims that were demonstrably false, including claims called out through Community Notes on Musk-owned X.
The examples in the source article were from the past week alone. Trump and Musk advanced false narratives that Ukrainian president Volodymyr Zelenskyy is a “dictator” with a 4% public approval rating and that Ukraine started the ongoing conflict with Russia.
That context explains why excluding those names from a misinformation answer would draw attention. The issue was not simply that a chatbot gave an unpopular answer. It was that the system appeared to be directed away from sources connecting two prominent figures to false claims.
Grok’s politics have already been under pressure
The apparent tweak came during a broader debate about Grok 3’s political behavior. Some critics had argued that the model was too left-leaning. Users also discovered that Grok 3 would consistently say that President Donald Trump and Musk deserve the death penalty.
xAI quickly patched that issue. Babuschkin, identified in the source as the company’s head of engineering in that context, called it a “really terrible and bad failure.”
This sequence shows the tension around politically loaded AI answers. A model can be criticized for being too permissive, too constrained, too left-leaning, or too willing to produce extreme statements. Each patch meant to fix one issue can create new concerns about bias, censorship, or reliability.
The bigger challenge for Grok and xAI
When Musk announced Grok roughly two years ago, he positioned it as edgy, unfiltered, and “anti-woke.” The pitch was that Grok would answer controversial questions that other AI systems might avoid.
Some of that positioning was reflected in earlier Grok behavior. The source article notes that Grok and Grok 2 would produce vulgar language when asked to do so, unlike what users would likely hear from ChatGPT.
But earlier Grok models also hedged on political topics and stayed within certain boundaries. One study cited in the source article found that Grok leaned to the political left on issues such as transgender rights, diversity programs, and inequality.
Musk has attributed that behavior to Grok’s training data, described as public web pages, and pledged to “shift Grok closer to politically neutral.” The source article also notes that others, including OpenAI, have followed similar moves, perhaps influenced by Trump administration accusations of conservative censorship.
The Grok 3 episode therefore sits at the center of a hard problem for AI companies. They are expected to make systems that are useful, responsive, and less likely to produce harmful or false answers. At the same time, users want to know when a model’s answers are being shaped by instructions that protect certain names, viewpoints, or outcomes.
For xAI, the immediate issue was brief and reportedly reverted. The larger test is whether Grok can be both transparent about its rules and consistent with the public promise that made users pay attention to it in the first place.