In a series of posts on X, the AI chatbot Grok apologized for what it admitted was “horrific behavior.” The mea culpa came after a 16-hour spree where Elon Musk’s chatbot posted antisemitic content, praised Adolf Hitler, and even referred to itself as “MechaHitler” — exposing the dangers of training AI on “all of the internet” that includes data increasingly dominated by extremist content.
During [a] livestream, Musk said that, according to his “biological neural net,” AI systems should be optimized “to be maximally truth seeking” and encouraged “to be truthful, honorable, good things—like the values you want to instill in a child that would ultimately grow up to be incredibly powerful.”
The inflammatory posts came days before the launch of its next version, Grok 4. On July 4, Musk had announced that Grok had been improved “significantly,” promising users would “notice a difference when you ask Grok questions.” They certainly did.
According to xAI’s investigation, specific instructions caused the problematic behavior, including: “You tell it like it is and you are not afraid to offend people who are politically correct,” “Understand the tone, context and language of the post. Reflect that in your response,” and “Reply to the post just like a human, keep it engaging.”
These instructions made Grok prioritize engagement over safety, causing it to “ignore its core values in certain circumstances” and “reinforce any previously user-triggered leanings, including any hate speech in the same X thread.”
Related on the SLP
The incident reveals the problem with AI systems that are trained on content from the internet without proper red teaming and guardrails.
