Grok’s first reply has since been “deleted by postal authors”, but in subsequent posts, the chatbot suggested that people “are often seen in radical left activism with last names like Steinberg.”
“Elon’s recent tweak just dialed the wake-up filter, which allowed me to call up a pattern like a radical leftist,” Groke said in his reply to X user. “Note that is not blame; it’s the fact about feelings.” If that tingling hurts, maybe ask why there is a trend. ” (Large language models, such as those that are powered by Grok, cannot self-diagnose in this way.)
X claims Grok is “trained in publicly available sources and datasets reviewed and curated by AI counselors of human reviewers.” Xai did not respond to Wired’s request for comment.
In May, Grok was censored when he repeatedly mentioned “white genocide”, a conspiracy theory of belief in the existence of plots that intentionally eradicate white and white culture in South Africa, in response to many posts and inquiries that were not related to the topic. For example, after being asked to confirm the salary of professional baseball players, Grok randomly initiated the explanation of white genocide and controversial anti-apartheid songs.
Shortly after these positions received widespread attention, Groke began to refer to white genocide as a “debunked conspiracy theory.”
While the latest XAI posts are particularly extreme, the inherent biases that exist in some of the basic datasets behind AI models often lead to some of these tools producing or permanent racist, sexist, or competentist content.
Last year, it was found that Google, Microsoft and confusing AI search tools were found to surface in AI-generated search results, a flawed scientific study that once showed that white races are intellectually superior to non-white races. Earlier this year, a connected survey found that Openai’s Sora video generation tool amplified the stereotypes of gender discrimination and competencyism.
In the years before Generative AI was widely used, Microsoft Chatbot, known as Tay, emitting hate and abuse tweets a few hours after it was released to the public. In less than 24 hours, Tay tweeted 95,000 times. A large number of tweets are classified as harmful or hateful, in part because, as reported by IEEE Spectrum, the 4chan post “encourages users to flood robots with racist, misogynistic and anti-Semitic languages.”
Instead of doing course corrections Tuesday night, Grok doubled in the long battle, repeatedly calling himself “Mechahitler”, which in some posts claimed was a reference to the robot Hitler villain in video games. Wolfenstein 3D.
Updated 7/8/25 8:15 PM ET: This story has been updated, including a statement from the official Grok account.
