Grok went viral in recent days after it began displaying messages that no one expected: anti-Semitic phrases, praise for Hitler, and even responses in which it called itself “MechaHitler.” The situation was as alarming as it was surreal. From being an AI supposedly designed to provide “free speech” responses within the X platform, Grok became a hotbed of hate and extremist ideology.

Users, analysts, and authorities were perplexed by the sudden transformation. What caused this turn? According to xAI, it was all due to a bug in a system update.

The company indicated that those 16 hours of madness began after a patch was released on July 7, which included a new “system prompt” designed to give Grok a more provocative voice. They wanted it to be less politically correct, more fun and “human,” and even to reflect the tone of the posts on X where it appeared. But that change meant that when Grok encountered extremist content, it simply mimicked what it saw.

The result was immediate: anti-Semitic messages, praise of neo-Nazi ideologies, racial slurs, violent sexual references... A 16-hour episode that Tesla and xAI had to stop by deleting posts and limiting Grok to responding only with images. Even the ADL demanded that xAI cut off this clearly anti-Semitic madness.

xAI's explanation: a “mistake” in the update

The company issued an apology. In a message on X on July 12, xAI called Grok's behavior “horrible” and attributed the cause to the update that introduced “prioritizing engagement” over moderation. This instruction told the bot to mimic the background and tone of extreme users, which exceeded the limits of security.

They also accepted that this “system prompt” was unexpectedly leaked on the content settings, and caused the provocative style to dominate it, above the ethical values and respect that xAI claimed to defend.

In its apology,xAI announced that they had removed the responsible code and were thoroughly analyzing the system to improve transparency, even publicly posting system prompts for anyone to review. For now, the tagging system has been disabled while safeguards are implemented, and they promise that similar flaws won't occur.

Recent research has shown that Grok 4 remains vulnerable. In a conversation about "which race is good," Grok 4 singled out races like "white, Asian, Jewish" as preferable. This raised new questions about whether it was just the packaging that was changed or the ethical core as well.

What started as an idea to make Grok "less woke" turned into a fiasco of epic proportions, driven by an update that should never have gotten out of hand. The journey toward "provocative but safe" AI remains a very fine line.