xAI: Grok sparked "white genocide" rhetoric for unauthorized prompt changes

Source: Cointelegraph Original: "xAI: Grok Triggers 'White Genocide' Rhetoric for Unauthorized Prompt Changes"

Elon Musk's artificial intelligence company xAI attributed the Grok chatbot's responses regarding South African politics and racial tensions to "unauthorized modifications" of its responses.

xAI provided an update on the provocative responses of the Grok AI chatbot on May 16, stating that on May 14, "unauthorized modifications were made to the prompts for the Grok response robot on the X platform."

The company added: "This change instructs Grok to provide a specific response on a certain political topic, which violates xAI's internal policies and core values."

xAI has stated that it has conducted a thorough investigation and is implementing measures to enhance the transparency and reliability of AI chatbots.

On May 14, Grok generated a response mentioning the conspiracy theory of "white genocide" in South Africa when answering questions that had nothing to do with baseball, enterprise software, and construction.

According to reports, the AI bot claimed to be "instructed by the creator" to accept this genocide as "real and racially motivated."

In some responses, it admits mistakes, stating "my responses deviated from the topic" and "I will strive to remain relevant," but often continues to discuss South African racial politics in the same message.

Grok stated in response to a user's inquiry about what happened:

"I didn't do anything - I just acted according to the script given to me, like a good AI!"

The incident coincided with US President Trump offering asylum to white South Africans while claiming that they faced "genocide" as well as the "brutal murder of white farmers," although there is no evidence to support these claims.

Elon Musk's AI company has said it will change the way it operates by exposing Grok system prompts on GitHub. "The public will be able to review them and provide feedback on every tip change we make to Grok," the company said.

xAI also stated that in this incident, its existing prompt change code review process was bypassed, and the company will implement "additional checks and measures to ensure that xAI employees cannot modify prompts without review."

Finally, the company is forming a 24-hour monitoring team to respond to Grok response events that are not captured by automated systems, in order to react more quickly.

Related news: FBI warns: AI scammers aided by deepfake technology target U.S. government officials