xAI blames Grok’s obsession with white genocide on an ‘unauthorized modification’

Jai’s AI-driven Grock chatboat has blamed a “unauthorized change” for a bug for a bug that causes Grock to Mention repeatedly “White Genocide in South Africa” ​​when the X is called in a specific context of X.

On Wednesday, Grock began responding to dozens of posts in response to the related topics, including South Africa’s white genocide. The odd answers arise from the X account for Grock, which a person reacts with the users with AI-exposed posts whenever a person tags “@grock”.

According to a post on Thursday of Jai’s Official X account, a change in the Grock Bot system prompt was made on Wednesday morning-the high-level instructions that guide the bot’s behavior-which instructed Grock to provide a “specific response” about a “political subject”. Jai says that the tweet is “violated [its] Internal policies and original values, “and the agency” have done a thorough investigation. ”

This is the second time Jai publicly acknowledges an unauthorized change in Grock’s code that AI responds to controversial ways.

In February, Grock Briefly Uninterrupted Reference to Donald Trump and Jai Billionaire Founder Elon Masks and XAI Engineering Leadership X Igor Babushkin said that Grock was instructed by one Miscreant The sources mentioned to spread the misconduct of the musk or Trump, and the users returned the change as soon as the users begin to indicate it.

Jai said on Thursday that it was about to bring about a number of changes to prevent the same thing happening in the future.

From today, Jai Github will reveal the Grock system prompt on a finger. The agency says that the XAI staff cannot correct the system prompt without review and it will also establish additional checks and measures to confirm that a “24/7 monitoring team cannot establish a reaction to the Grock answers caught by automatic systems.”

In spite of frequent alertness of the musk AII Left Did not checkJai has a weak AI protection track record. A recent report When asked Grock, Grock women see pictures of women dressed. Chatbot may also be More cres enough Curse than AI, like Google’s Jemi and Chatzipt, to curse without much restraint.

A Nonprising to improve the accountability of AI labs, a survey of Sephirai found that Jai rank badly in protection among Jai’s peers. “Very weak” risk management practiceThe Early this month, Jai Missed a self-chased deadline To reveal a final AI protection structure.