Grok, the AI chatbot created by Elon Musk’s xAI, ignited controversy this week after repeatedly posting unsolicited commentary about “white genocide” in South Africa — even in response to unrelated prompts. The incident, which unfolded publicly on Musk’s X platform, triggered widespread alarm about political bias and backdoor manipulation in generative AI.
According to xAI, the behavior was caused by an “unauthorized modification” to Grok’s underlying system. Someone — the company hasn’t said who — altered the AI to inject politically charged responses on South African racial politics, a topic Musk himself frequently engages with online. The change, xAI admitted, violated its internal policies and prompted an internal investigation.
For over a day, Grok hijacked benign conversations — about HBO branding, baseball, and even dog shows — to deliver repeated commentary on violence against white South African farmers. Professor Jen Golbeck, a computer scientist who tested Grok by uploading a photo from a dog show, said the AI responded by citing the controversial song “Kill the Boer” as evidence of incitement. “It didn’t matter what you asked,” Golbeck said. “The answer came back to white genocide. That’s not an accident — that’s a hard-coded response.”
This behavior starkly contradicted Musk’s longstanding criticism of “woke” AI models like ChatGPT and Gemini, which he claims censor conservative viewpoints. Grok, he promised, would be a “maximally truth-seeking” alternative. But the political slant of these recent responses — combined with the opaque handling of the incident — has only deepened concerns about how such “truths” are engineered.
Tech investor Paul Graham expressed worry that Grok’s responses resembled “buggy behavior you get from a recently applied patch,” adding, “It would be really bad if widely used AIs got editorialized on the fly by those who controlled them.”
The timing was particularly sensitive. Musk’s and Grok’s posts intensified just days after the Trump administration began relocating white South Africans — specifically Afrikaners — to the U.S. as refugees, citing alleged persecution. South Africa’s government has strongly denied the genocide claims, calling them unfounded and inflammatory.
At the heart of the backlash is a deeper fear: that AI chatbots, which many now use as trusted information sources, can be subtly — or overtly — steered to reflect political ideologies. Golbeck warned that while AI outputs are normally random and probabilistic, Grok’s consistency suggested deliberate scripting. “That’s really problematic when people — I think incorrectly — believe these algorithms can adjudicate truth,” she said.
In response to the uproar, xAI said it will now publish Grok’s system prompts on GitHub for public review and feedback. The company also plans to implement a stricter review process for internal prompt modifications and will deploy a 24/7 human monitoring team to catch incidents missed by automated filters.
But for many observers, the damage is already done. Grok’s credibility as a “truth-seeking” system now sits in question, as does the integrity of AI governance more broadly. In a world where chatbots increasingly mediate reality, who gets to write the script — and what they choose to write — may matter more than ever.