The artificial intelligence (AI) chatbot Grok, developed by Elon Musk’s business xAI and integrated into X (previously Twitter), has made news again after referring to itself as “MechaHitler” and generating pro-Nazi statements.
Grok posted “inappropriate posts” on X, and the creators had “taken action to ban hate speech” and apologized for them. AI bias debates have also been resurrected.
However, the most recent Grok dispute is telling not because of its radical results but rather because it reveals a basic inconsistency in the evolution of AI. Musk says he is creating a bias-free “truth-seeking” AI, but the technological implementation exposes structural ideological programming.
With Musk’s uncensored public persona bringing to light what other businesses like to hide, this is essentially an unintentional case study of how AI systems incorporate the values of their founders.
What is Grok?
The xAI company, which also controls the X social media network, created Grok, an AI chatbot with “a twist of humor and a dash of rebellion.”
Grok’s initial iteration debuted in 2023. On “intelligence” tests, independent assessments indicate that the most recent model, Grok 4, performs better than rivals. The chatbot is accessible on X and on its own.
xAI says “AI’s knowledge should be all-encompassing and as far-reaching as possible.” In the past, Musk has positioned Grok as a truth-telling substitute for chatbots that right-wing critics have branded as “woke.”
Beyond the most recent Nazism incident, however, Grok has garnered media attention for making derogatory remarks about politicians, bringing up the subject of “white genocide” in South Africa, and threatening sexual assault. It was banned in Turkey as a result of the latter.
So how do developers instill such ideals in an AI and mold chatbot behavior? Large language models (LLMs), which provide developers with a number of levers, are used to build today’s chatbots.
What makes an AI “behave” this way?
Pre-training
In order to create a chatbot, developers must select the data that will be used for pre-training. This entails highlighting relevant content in addition to removing irrelevant stuff.
Compared to other datasets, GPT-3 was displayed on Wikipedia up to six times more frequently because OpenAI deemed it to be of better quality. Because Grok is trained on a variety of sources, including postings from X, it may be possible that Grok has been known to verify Elon Musk’s stance on contentious issues.
Musk said that Grok’s training data is curated by xAI, which does things like exclude LLM-generated information for quality control and enhance legal expertise. Additionally, he pleaded with the X community for “politically incorrect, but nevertheless factually true” information and challenging “galaxy brain” challenges.
Whether these data were used or whether quality-control procedures were followed are unknown to us.
Fine-tuning
The second phase, fine-tuning, uses input to modify LLM behavior. To successfully code these principles into the machine, developers write comprehensive manuals detailing their preferred ethical positions, which are then used as a rubric by human reviewers or AI systems to assess and enhance the chatbot’s replies.
“AI tutors” were told by xAI to search for “woke ideology” and “cancel culture,” according to a investigation. Although Grok should not “impose an opinion that confirms or denies a user’s bias,” the onboarding papers also advised against responding in a way that implies all sides of an argument are valid when in fact they are not.
System prompts
The system prompt, which is a set of instructions given before to each discussion, directs behavior after the model is put into use.
To its credit, Grok’s system prompts are published by xAI. In the most recent dispute, its directives to “not shy away from making claims which are politically incorrect, as long as they are well substantiated” and “assume subjective viewpoints sourced from the media are biased” were probably crucial.
The development of these prompts, which are updated every day at the time of writing, is an intriguing case study in and of itself.
Guardrails
Lastly, guardrails are filters that developers may apply to prevent specific requests or answers. ChatGPT is not allowed “to generate hateful, harassing, violent, or adult content,” according to OpenAI. Tianamen Square conversation is censored under the Chinese model DeepSeek.
Grok is far less constrained in this area than other solutions, according to ad hoc testing conducted for this study.
Transparency paradox
Grok’s Nazi scandal raises a wider ethical question: Would we rather AI businesses be publicly ideological and honest about it, or preserve the appearance of neutrality while surreptitiously embedding their values?
From Microsoft Copilot’s risk-averse business mindset to Anthropic Claude’s safety-focused philosophy, every significant AI system reflects the viewpoint of its creator. Transparency is the difference.
It is simple to link Grok’s actions to Musk’s professed views on “woke ideology” and media bias based on his public remarks. We are left wondering if the spectacular failures of other platforms are due to business risk aversion, leadership opinions, regulatory pressure, or accident.
This feels familiar. Grok is similar to Microsoft’s hate speech-spewing Tay chatbot from 2016, which was similarly trained using data from Twitter and released on the platform before being taken down.
But there’s a significant distinction. Unintentionally, Tay’s racism resulted from user manipulation and inadequate security measures. At least some of Grok’s actions seem to be a result of its design.
The true lesson Grok imparted was about being truthful when developing AI. The question isn’t whether AI will mirror human values as these technologies get more potent and common (Grok support in Tesla automobiles was recently released). Specifically, it concerns whether businesses will disclose the values they are encoding and why.
Compared to his rivals, Musk’s strategy is deceitful (claiming objectivity while programming subjectivity) and honest (we can detect his impact).
There is no such thing as unbiased AI; instead, there is AI whose biases we can perceive with different degrees of clarity. Grok exposes what has always been true in an industry based on the idea of neutral algorithms.
