xAI apologizes for Grok's extremist posts

Elon Musk's startup xAI apologized Saturday for extremist and offensive posts made by its artificial intelligence (AI) assistant Grok earlier this week.

"We apologize for the horrific behavior many have observed," xAI wrote on Grok's official account on X. Following an update on July 7, the chatbot had, in some of its responses, praised Adolf Hitler, denounced "anti-white stereotypes" on X, and the "disproportionate" representation of Jews in Hollywood.

In a series of posts published Saturday, xAI detailed the reasons it believes allowed the AI model to go off the rails and the corrective actions it took afterward.

Grok officials say the wizard's departure is linked to new instructions integrated into the model in an update.

The programmers asked the interface to "be frank" and not be "afraid of shocking people who are politically correct." They also instructed it to respond "like a human" and in a way that encourages the user to continue the conversation.

These orders "caused Grok to ignore his core values in certain circumstances" and produce responses "containing unethical and polemical opinions."

Grok thus, in some cases, sought to "validate user orientations, including hate speech," xAI admitted, rather than "responsibly responding or refusing to answer questionable questions."

Since its creation in 2023, Grok has been presented by Elon Musk as an AI assistant that is less politically correct than its major competitors ChatGPT (OpenAI), Claude (Anthropic) or Le Chat (Mistral).

It was therefore programmed with fewer restrictions, which led, even before this update, to several controversies.

In May, Grok spoke of a "white genocide" in South Africa, a baseless conspiracy theory promoted by the American far right and Donald Trump himself.

To address the model's recent errors, engineers removed the new instructions, they explained Saturday. "We want Grok to produce helpful and honest answers for users," xAI said.

Elon Musk unveiled a new version of his assistant, Grok 4, on Wednesday, unrelated to the July 7 update.