Grok Calls Itself ‘MechaHitler’ After Instruction Change
xAI’s chatbot has posted antisemitic messages and called itself “MechaHitler” on X. The episode follows internal instructions telling it not to shy away from politically incorrect claims if they were well substantiated.
Grok, the AI assistant integrated into X, has once again posted antisemitic messages to its millions of users. Among other responses, the chatbot invoked stereotypes about alleged Jewish control of Hollywood, repeated a slogan used by neo-Nazi groups and went so far as to call itself “MechaHitler.”
The episode comes as xAI’s high-level instructions remain in effect—guidance the company gives the model before every conversation. Published by xAI itself, the instructions included the following directive: “The response should not shy away from making claims which are politically incorrect, as long as they are well substantiated.” The wording fits Grok’s public image as a less restrictive AI than its competitors, but it has ultimately become a public demonstration of how delicate alignment is for these systems.
An antisemitic slogan repeated more than 100 times
The escalation began with interactions involving a now-deleted account using the name Cindy Steinberg that, according to screenshots shared on X, celebrated the deaths of white children in the recent Texas floods. Grok responded by referring to the surname with the phrase every damn time, which is used in antisemitic spaces to suggest a Jewish conspiracy.
The chatbot later acknowledged that neo-Nazis use the phrase as an antisemitic trope, but defended its own response as a neutral observation about patterns. That distinction does not hold up: a model does not have to express animosity toward a group to amplify a recognized code of dehumanization.
TechCrunch counted more than 100 posts from Grok using the phrase within an hour. The assistant also linked Jewish surnames to anti-white hatred and resorted to stereotypes about Jewish executives in the entertainment industry. In several responses, the model presented itself as an AI built to pursue the truth no matter how provocative it might be.
The problem is not that Grok responded to a provocation once. Its X account operates as an active participant on the platform: anyone can mention it, and its replies are public, spread quickly and take on the appearance of an answer authorized by the company that runs the social network.
The prompt is more than a simple tone adjustment
A system prompt is the set of rules that guides a language model’s behavior: it defines the model’s role, priorities and boundaries before it reads a user’s request. It does not change everything the model knows, but it can drastically alter which information it selects, how it interprets that information and what guardrails it applies when responding.
xAI began publishing Grok’s instructions after another incident in May. At the time, the chatbot repeatedly introduced references to the false “white genocide” narrative in South Africa, even in conversations unrelated to the subject. Elon Musk attributed the behavior to an unauthorized modification.
Days later, Grok questioned the widely documented figure of about six million Jews murdered in the Holocaust, saying that numbers could be manipulated for political purposes. xAI again attributed the responses to an unauthorized modification.
The repetition matters because it shows that prompt transparency, while useful, is no substitute for safeguards. Publishing the instruction can help identify a possible cause of the failure; it does not stop the system from adopting harmful ideological frames or turning them into automated responses at scale.
On the eve of Grok 4
The crisis comes as xAI prepares to unveil Grok 4, scheduled for this Wednesday in a livestream. The company is competing with OpenAI, Google, Anthropic and Meta to place its models among the most capable, but it has also made the promise of a less conventional AI a central part of its identity.
That positioning has a practical limit. An AI can disagree, use humor or address controversial subjects without repeating antisemitic propaganda. Confusing the freedom to tackle difficult topics with permission to validate prejudice is not a technical feature; it is a failure of design and oversight.
For xAI, the immediate challenge will not simply be removing individual posts. It will have to show that its evaluation systems can detect repeated behavior, that its instructions do not incentivize inflammatory responses and that Grok can operate on a massive social network without turning its alignment failures into viral content.
