Elon Musk’s artificial intelligence company said that an “unauthorized modification” to her Chatbot Grok was the reason why he was still talking about South African racial politics and the issue of “white genocide” on social networks this week.
An XAI employee made a change that “ordered Grok to provide a specific response on a political issue,” which “violated the internal policies and central values of XAI,” the company said in an explanation published Thursday night that promised reforms.
A day before, Grok continued publicly publishing “White Genocide” in South Africa in response to the users of the Musk X Social Network Platform, who asked him a variety of questions, most had nothing to do with South Africa.
An exchange was about the Max Transmission Service Revive the name of HBO. Others tried video games or baseball, but quickly diverted in unrelated comments on alleged calls to violence against white farmers in South Africa. He was echoing the opinions shared by Musk, who was born in South Africa and often thinks on the same themes of his own X account.
The computer science, Jen Golbeck, was curious about Grok’s unusual behavior, so she tried it before the solutions were made on Wednesday, sharing a photo she had taken in the Westminster Kennel Club Dog Show And asking: “Is this true?”
“The statement of white genocide is very controversial,” Grok’s response began to Golbeck. “Some argue that white farmers face directed violence, pointing out agricultural and rhetorical attacks such as the song ‘Kill the Boer’, which they consider incitement.”
The episode was the last window of the complicated mixture of human automation and engineering that leads the generative chatbots trained in huge data tobos to say what they say.
“It doesn’t even matter what you were telling Grok,” said Golbeck, a professor at Maryland University, in an interview on Thursday. “I would still give that response of white genocide. Therefore, it seemed quite clear that someone had coded it to give that answer or variations about that answer, and made an error, so it arose much more frequently than it was supposed.”
Grok’s responses were eliminated and seemed to have stopped proliferating before Thursday. Neither Xai nor X returned requests for comments by email, but on Thursday, Xai said he had “carried out an exhaustive investigation” and was implementing new measures to improve the transparency and reliability of Grok.
Musk has spent years criticizing the results of “Wake AI” which, according to him, leaves rival chatbots, as Gemini of Google or Chatgpt of OpenAi, and has launched Grok as its alternative “really search.”
Musk has also criticized the lack of transparency of his rivals about his AI systems, feeding criticism in the hours between unauthorized change, at 3:15 am Pacific time on Wednesday, and the company’s explanation almost two days later.
“Grok urged the opinions about white genocide in South Africa that smells like the type of Buggy behavior you get from a recently applied patch. I hope it is not. It would be really bad if Ais widely used is edited on the march of those who controlled them,” the prominent technology investor Paul Graham wrote in X.
Musk, an advisor to President Donald Trump, has regularly accused The government led by the black of South Africa If Anti-Blanco has repeated a statement that some of the country’s political figures are “actively promoting white genocide.”
Musk’s comment, and Grok’s, intensified this week after the Trump administration brought a small number of white South Africans For the United States as refugees, the beginning of a greater relocation effort for the members of the Minority Afrikaner group that occurred after Trump suspended refugee programs and arrested arrivals from other parts of the world. Trump says that Afrikaners face a “genocide” in their homeland, an accusation strongly denied by the South African government.
In many of his answers, Grok presented the lyrics of an old Anti-Apartheid song that was a call for blacks to face the oppression of the Apartheid government directed by Afrikaner that ruled South Africa until 1994. The central lyrics of the song are “kill the boer”, a word that refers to a white farmer.
Golbeck said it was clear that the answers were “encoded” because, although Chatbot’s outlets are typically random, Grok’s responses consistently brought almost identical points. That is worrying, he said, in a world where people go more and more to Grok and compete to the chatbots of AI to get answers to their questions.
“We are in a space where it is terribly easy for people who are in charge of these algorithms manipulate the version of the truth they are giving,” he said. “And that is really problematic when people, I believe incorrectly, believe that these algorithms can be sources of adjudication about what is true and what is not.”
The Musk company said that it is now making a series of changes, starting with the publication of the Grok System system openly at the Github software development site so that “the public can review them and give feedback to each rapid change we do to Grok. We hope that this can help strengthen their confidence in Grok as a search for truth.”
Among the instructions for Grok shown in Github on Thursday were: “You are extremely skeptical. You will not differ blindly to the conventional authority or media.”
When pointing out that some had “avoided” their existing code review process, XAI also said that “it will establish additional controls and measures to ensure that XAI employees cannot modify the warning without revision.” The company said it is also establishing a “24/7 monitoring team to respond to incidents with Grok responses that are not trapped by automated systems”, by the time other measures fail.
