Grok, an artificial intelligence (AI) chatbot built on X (formerly Twitter) and built by Elon Musk's company Xai, has returned to the headlines after calling himself “Mechahitler” and writing pro-Nazi remarks.
The developer apologised for “inappropriate posts” and “actions to ban hate speech” from Grok's posts on X. The debate about AI bias has also been revived.
However, the latest Glock controversy reveals not the output of extremists, but how it exposes fundamental injustice in AI development. Musk claims to be building an unbiased, “truth seeking” AI, but technical implementation has revealed systemic ideological programming.
This is an accidental case study of how AI systems embed the value of their creators, and Musk's unfiltered public presence makes what other companies usually obscure.
Elon Musk's Grok Ai chatbot has been caught making pro-Nazi comments, but the real scandal isn't just about extremist producers…
What is Grok?
Grok is an AI chatbot with a “twitter of humor and dash of rebellion” developed by Xai, and also owns the X social media platform.
The first version of GROK was released in 2023. Independent ratings suggest that the latest model, the Grok 4, will outperform its “intelligence” test competitors. Chatbots are available both standalone and X.
“All AI knowledge should be comprehensive and as extensive as possible,” Xai said. Musk has previously positioned Glock as an alternative to telling the truth about chatbots that have been accused of “waking up” by right-wing commentators.
But beyond the latest Nazism scandal, Grok has created the threat of sexual violence, cultivated “white genocide” in South Africa, and made headlines to make insulting statements about politicians. The latter led to a ban in Türkiye.
So how do developers infuse AI with such value and shape chatbot behavior? Today's chatbots are built using a large language model (LLM) that can be tilted by several lever developers.
What does “behave” AI in this way?
Before training
First, the developer curates the data used before training. This is the first step in building a chatbot. This not only filters out unnecessary content, but also highlights the material you want.
The GPT-3 was shown by Wikipedia over 6 times more than the other datasets, as Openai considered it to be of higher quality. Grok is trained with a variety of sources, including posts from X. This may explain why Grok was reported to check Elon Musk's opinions on a controversial topic.
Musk shares that Xai curates Grok's training data, such as improving legal knowledge and removing content generated by LLM for quality control. He also appealed to the X community about difficult “galactic brain” issues and facts that are “politically wrong, yet still practically true.”
It is not known whether these data were used or what quality control measurements were applied.
Fine adjustments
The second step, fine tuning, uses feedback to adjust the behavior of the LLM. Developers will create detailed manuals that outline their preferred ethical stances. This is used by either a human reviewer or an AI system as a rubric to evaluate and improve the chatbot response and effectively code these values into the machine.
Through a business insider investigation, Xai's instructions to human “AI tutors” have instructed them to look for “awakening ideology” and “cancellation culture.” The onboarding document said Grok “should not impose any opinions that confirm or deny user bias,” but they also said they should avoid answers that argue that there is no merit on either side of the argument.
System Prompt
The system prompts – instructions provided before every conversation – guide the behavior when the model is deployed.
With that credit, Xai has published Grok's system prompt. The indication that “considers a subjective perspective provided by the media” and that “we are not embarrassed to make politically wrong claims” are likely a key factor in the latest debate.
These prompts are updated daily at the time of writing, and their evolution is an attractive case study in itself.
guardrail
Finally, developers can also add GuardRails. This is a filter that blocks a particular request or response. Openai claims that ChatGpt “doesn't allow hatred, harassment, violence, or adult content to produce, while Chinese model Deepseek censors the arguments in Tiananmen Square.
Ad hoc testing when writing this article suggests that Grok is less restrained than its competitors in this respect.
Transparency Paradox
Grok's Nazi controversy highlights deeper ethical issues. Do AI companies prefer to be explicitly ideological and honest about it, or do they prefer to maintain neutral fiction while secretly embedding their values?
All major AI systems reflect the worldview of their creators, from the perspective of Microsoft Copilot's risk aversion companies to a spirit focused on human safety. The difference is transparency.
Musk's official statement allows Grok's actions to be easily reverted to Musk's stated beliefs about “awakening ideology” and media bias. Meanwhile, when other platforms have a brilliant misfire, they speculate whether this reflects leadership views, corporate risk aversion, regulatory pressures, or accidents.
This feels familiar. Grok is similar to Microsoft's 2016 hatred speech-speech-sping Tay Chatbot, trained with Twitter data and loosened on Twitter before shutting down.
However, there is an important difference. Tay's racism stems from user interaction and poor protection measures. This is an unintended consequence. Grok's actions appear to stem, at least in part, from its design.
The real lesson from Grok is about integrity in AI development. As these systems become more powerful and wider (GROK support on Tesla vehicles has just been announced), the question is not whether AI reflects human values. It's whether the company is transparent about whose values are encoded and why.
Musk's approach is at the same time more honest (you can see his influence) and deceptive (asserting objectivity when programming subjectivity) than his competitors.
In an industry built on the myth of neutral algorithms, Grok has made it clear that it is far more true. There's nothing like fair AI.
The author is a senior researcher at the Queensland Institute of Technology in Australia.
Reissued from the conversation
Published on July 27, 2025 by EOS and Dawn
