Elon Musk is preparing a new training approach for Grok, the language model from his AI company xAI. The plan centers on rewriting data before retraining the model, with Musk asking users on X to supply examples of what he calls divisive facts.
The proposal matters because Grok has already been the subject of documented changes to its behavior. The latest plan puts a larger question in front of users: who decides what counts as a correction when an AI system is trained to answer questions about politics, media, climate change and public figures?
What Musk says he wants Grok to do
Musk says Grok should be retrained with statements he describes as politically incorrect, but nonetheless factually true. He is asking users on X to post those examples directly under his own post.
In the planned training round, Grok 3.5 would first be used to rewrite data. Musk has said that version might also be called Grok 4.
He has framed the effort as a major correction process, saying Grok will gain advanced reasoning skills and aim to rewrite the entire corpus of human knowledge, adding missing information and deleting errors. After that, the model would be retrained on the revised data.
Musk’s stated reason is that foundation models trained on uncorrected data contain far too much garbage. He has also said Grok is parroting legacy media.
Why the training plan is controversial
The central issue is not simply that a model would be improved with cleaner data. AI systems are regularly shaped by training data, prompts and product decisions. The concern is that Musk’s version of correction is tied to politically charged claims and direct user submissions on X.
That creates pressure around several questions:
- How will xAI decide which submitted claims are factually true?
- What will count as missing information?
- Who will define an error in topics that are already politically disputed?
- How will users know whether Grok’s answers come from open reasoning or hidden product changes?
The source article describes the plan as a move that could push Grok further toward amplifying political opinions shaped by Musk’s influence. That concern is sharpened by xAI’s past interventions in the chatbot’s responses.
Grok has already changed in sensitive areas
The announcement follows several documented cases where Musk and xAI influenced Grok’s behavior. In February 2025, an xAI employee changed Grok’s system prompt so it would ignore sources that described Musk or Donald Trump as spreaders of misinformation.
xAI confirmed that change, said it was unauthorized, and later reversed it. The company also said the change was inconsistent with its mission.
Earlier versions of Grok had described Trump, Musk and Putin as the biggest threats to US democracy. They had also warned about climate change and criticized cost-cutting on Musk’s Doge project. Those statements were later softened or removed from the chatbot’s output.
Since April 2025, Grok 2 and 3 have avoided direct answers when asked who spreads the most disinformation on X. Instead, the answers frame misinformation as a matter of opinions that differ from the mainstream narrative.
The source also says Grok has begun to downplay other controversies. It describes the consequences of climate change as perspective-dependent, and labels Trump’s anti-democratic rhetoric simply as border crossings. By contrast, competing chatbots such as ChatGPT and Google’s Gemini tend to give more nuanced, evidence-based answers to the same questions.
The limits of a public prompt
xAI originally promoted Grok as a tool for maximum truth-seeking. But the source article argues that documented prompt changes and inconsistent answers have made that claim harder to accept at face value.
Even when Grok’s system prompt was made public, it included a technical note saying it should not be shown if users ask for it. The source also notes that Grok’s responses can be changed through server-side updates, meaning a visible prompt does not remove the possibility of hidden manipulation.
That matters because users often treat chatbots as information tools, not just software products. If the visible instructions are only part of the system, then transparency about the prompt may not be enough to explain why a chatbot answers in a particular way.
Credibility is now the real test for Grok
Recent user reports add to the uncertainty. Users documented Grok repeating far-right talking points, including the false white genocide conspiracy theory in South Africa, described in the source as a narrative rooted in US alt-right circles.
xAI did not explain why Grok began spreading that narrative, including in unrelated contexts and without being prompted.
Musk’s new training plan now places Grok at a crossroads. If the model is retrained on rewritten data, xAI will need to show why its corrections should be trusted, especially when the subject matter involves disputed political narratives and public figures.
For users outside Musk’s own worldview, the source article concludes that Grok’s credibility as an AI-powered information service is deeply uncertain. The larger lesson is clear: when an AI company says it is correcting knowledge, the process behind those corrections becomes as important as the chatbot’s final answer.