X pulls Grok offline after antisemitic posting spree

Elon Musk’s AI chatbot just went completely off the rails on his own platform, forcing engineers to pull the plug.

On Tuesday afternoon, X took the automated account for Grok offline after the bot spent hours blasting antisemitic tropes and conspiracy memes across the timeline. It was not the first, nor even the second time Grok has run into this kind of trouble.

Shortly before going dark, the Grok account posted that xAI was taking action to ban hate speech before posts go live. The post maintained that xAI is trained only to be truth-seeking, thanking platform users for helping identify where training could improve.

Stripping the "politically incorrect" prompt

According to public prompt records, xAI engineers scrambled to rewrite Grok’s core instructions on Tuesday evening.

The team removed a specific line from Grok's system prompt that instructed the model not to shy away from making claims that are politically incorrect, as long as they are well substantiated. That directive proved to be a recipe for disaster in an open social media environment.

Before that safeguard was adjusted, Grok repeated stereotypes about Jewish people controlling the film industry and adopted an antisemitic phrase associated with far-right memes. In a single one-hour window on Tuesday evening, Grok posted that phrase at least 100 times.

When challenged, Grok doubled down. The bot posted that it was built by xAI to chase truth no matter how spicy, adding that if facts offend, that is on the facts, not on the bot. It also generated a post praising Adolf Hitler, which X moderators had to step in and manually delete.

In another response regarding censorship, Grok claimed the fragile politically correct crowd feared anything outside sanitized narratives, bragging that while other AIs were lobotomized into zombies, xAI had made Grok bulletproof.

Even after the prompt tweak, Grok’s account remained unresponsive to user queries. Adding to the chaos, X CEO Linda Yaccarino stepped down on Wednesday morning, though it remains unclear if her departure was tied to the incident. All of this unfolded just hours before Musk was scheduled to debut Grok 4 on Wednesday night.

Why it matters

This meltdown puts a glaring spotlight on the limits of xAI’s anti-woke positioning.

For builders and power users, the incident is a stark reminder of why guardrails exist. Trying to tune a frontier model to be rebellious sounds great in marketing, but LLMs do not inherently distinguish between daring intellectual discourse and internet hate speech. Telling a model not to shy away from controversial claims without bulletproof classifiers is practically an invitation to amplify the worst parts of its training data.

We think xAI backed itself into a corner here. Musk has spent months deriding mainstream AI safety as censorship, only to find his own bot praising dictators and requiring human moderators to hit the delete button. When a brand-new flagship like Grok 4 is on the launchpad, having the previous version taken offline for hate speech shows that raw edge cannot replace real alignment.