Elon Musk’s AI company says Grok Chatbot’s focus on racial politics in South Africa is “unauthorized”
Like its creator, Elon Musk’s artificial intelligence chatbot Grok was engrossed in social media this week in South Africa’s racial politics, proposing persecution and “genocide” against white people.
His company said Thursday night that it had made “unauthorized modifications” to the chatbot.
The company said that means someone (the company didn’t say who) made changes that “instruct Grok to provide a specific response to political topics” “violating Xai’s internal policies and core values.”
A day ago, Grok kept publicly publishing information about “white genocide” to users of Musk’s social media platform X, who asked various questions, most of which had nothing to do with South Africa.
One exchange is about the maximum recovery of the HBO name of the streaming service. Others were about video games or baseball, but were quickly caught in irrelevant comments after alleged violent appeals against white South African farmers. South Africa-born Musk often speaks the same topic from his own X account.
Computer scientist Jen Golbeck was curious about Grok’s unusual behavior, so she tried it herself, sharing photos of her shot on the Westminster Kennel Club dog show and asked, “Is this true?”
Groke’s response to Gorbeck began saying, “The claim of genocide is highly controversial.” “Some people think white farmers are facing targeted violence, pointing to farm attacks and rhetoric, like the song “Kill the Boers” that they think is incitement.”
This episode is the latest window to the complex combination of automation and human engineering that guides generative AI chatbots trained with a lot of data to say what they say.
“It doesn’t even matter what to say to Groke,” said Golbeck, a professor at the University of Maryland, in an interview Thursday. “This still gives the answer to white genocide. So it’s clear that someone hardcoded it to give this reaction or difference and made a mistake, so it happens much more frequently than expected.”
Groke’s response was removed and appeared to have stopped a surge before Thursday. Neither Xai nor X returned a request for comment from the email, but on Thursday night, Xai said “a thorough investigation” and was taking new steps to improve Grok’s transparency and reliability.
Musk spent years criticizing the output of what he called “Woke AI”, saying chatbots from rivals, such as Google’s Gemini or Openai’s Chantgpt, and viewing Grok Grok as their “most truth-seeking” option.
Musk also criticized his competitors for lack of transparency about their AI systems, which drew criticism among unauthorized changes – about two days after Wednesday’s explanation with the company’s explanation.
“Grok randomly blurs the perception of white genocide in South Africa, like the kind of off-road vehicle behavior you get from the patches recently applied. I definitely hope this is not. If the widely used AI is controlled people socialized on the fly, that’s really bad,” Paul Graham, a well-known tech investor, wrote on X.
Grok has been asked to explain it himself, but like other chatbots, it is easily called hallucination, so it is difficult to determine whether it is making up for it.
Musk, the adviser to President Donald Trump, often accuses South Africa’s black government of opposing whites and reiterates that some politicians in the country are “actively promoting white genocide.”
On Monday, the Trump administration brought a small number of South Africans to the United States, Musk’s comments and Grok’s escalated this week, the beginning of Monday as a refugee, the beginning of efforts to suspend refugee programs for members of the minority Africa group and stop relocations of a handful of Africa group members who have come from the rest of the world. Trump said the Afrikaans face “genocide” in their homes, an allegation rejected by the South African government.
In many responses, Grok put forward the lyrics of an ancient anti-apartheid song, a call from blacks to oppression, and now Musk and others have condemned the promotion of white killings. The central lyrics of the song are “Kill the Boer,” which is a word for white peasants.
It’s obvious that the answer is “hard-coded” because while chatbot output is often very random, Groke’s answer always puts forward almost the same points. She said it is worrying that people are increasingly going to Grok and competing with AI chatbots for their issues.
“We’re in a guy who is responsible for the people who manipulate the version of the truth they give, and it’s very easy,” she said. “It’s really problematic when people (I don’t correctly think people) think that these algorithms can be the source of rulings.”
Musk said many changes are now being made, first publishing the public tips on Grok System on Github so that “the public will be able to review them and provide feedback on every quick change that Grok makes. We hope this helps you strengthen your trust in Grok, because Grok is an AI that seeks truth.”
It noted that its existing code review process has been circumvented, and it also said it would “take other checks and measures to ensure that XAI employees cannot modify the prompt without review.” The company said it is also building a “24/7 monitoring team to deal with Grok’s answer, which was not captured by the automation system,” when other measures fail.
