Grok kept talking about 'white genocide' due to an 'unauthorized modification'

In the early hours of May 14, xAI's chatbot Grok repeatedly gave X users responses that referred to claims about a "white genocide" in South Africa even if their inquiries had nothing to do about the subject. Now, in a statement posted on the social network, Elon Musk's AI company has explained that "an unauthorized modification" to Grok's prompt on X caused it to "provide a specific response on a political topic." It didn't say what had happened to the personnel involved in rolling out the rogue update. But it added that the modification violated its "internal policies and core values" and that it has conducted a thorough investigation about the incident. 

The website's various users had posted several instances wherein Grok included references to the controversial claims that white South African farmers are facing racial discrimination and land seizures in their country. Their questions? Well, in one tweet, someone asked how many times HBO has changed its streaming service's name. In another, the user asked a baseball player's salary history. In yet another one, someone asked for more information about a WWE match. CNBC was able to replicate the chatbot's responses with white genocide references. When the news website asked if it was specifically programmed to promote "white genocide," Grok said that it wasn't and that its "purpose is to provide factual, helpful, and safe responses based on reason and evidence."

Before xAI issued a response, OpenAI chief Sam Altman posted a snarky response on X. "I’m sure xAI will provide a full and transparent explanation soon," he wrote, and then mimicked Grok's responses by segueing into talking about white genocide. xAI said that from now on, it will be publishing its system prompts on GitHub so that the public can give feedback on every alteration. The company also said that it will put additional checks and measures to ensure xAI employees can't modify Grok's prompt without a review. Whoever edited it recently was able to circumvent its current review process in this case. In addition, the company said it's putting together a team that can monitor incidents related to Grok's answers not caught by automated systems 24/7.

As TechCrunch has noted, this isn't the first time xAI had blamed a contentious Grok behavior to an unauthorized change. Back in February, the chatbot briefly censored sources that talked about how Musk and President Donald Trump are spreading misinformation. xAI co-founder Igor Babuschkin said at the time that a rogue employee had pushed an unapproved modification to Grok's prompt.

This article originally appeared on Engadget at https://www.engadget.com/ai/grok-kept-talking-about-white-genocide-due-to-an-unauthorized-modification-120044119.html?src=rss https://www.engadget.com/ai/grok-kept-talking-about-white-genocide-due-to-an-unauthorized-modification-120044119.html?src=rss