Seattle - Microsoft on Monday said its cloud servers will now host Grok from Elon Musk's xAI, days after the chatbot went off the rails with talk of "white genocide" in South Africa.
Musk told an event hosted by Microsoft that his company's models "aspire to truth with minimal error," adding that "there's always going to be some mistakes that are made."
The Grok chatbot last week ignited controversy by answering unrelated user prompts with unbacked right-wing propaganda about purported oppression of white South Africans. In a recorded conversation with Microsoft CEO Satya Nadella, Musk said that xAI would always acknowledge mistakes with its Grok AI models. "It's incredibly important for AI models to be grounded in reality," the Tesla tycoon said. Generative AI models are often pre-programmed by engineers -- through things known as system prompts -- to give or avoid specific responses or convey certain moods or styles, no matter the input given by the user. Most recently, the latest model from industry leader OpenAI was found to be generating overly sycophantic responses, and the company quickly said it would make changes to remove the bug.
The answers provided by Grok drew alarm as they reflected a conspiracy theory often shared on social media by Musk, who was born in South Africa. The company did not identify who made the code change, but said an "unauthorized modification" directed Grok to provide a specific response that "violated xAI's internal policies and core values."
Faced with criticism, the startup said it was implementing measures to make Grok's system prompts public, change its review processes and put in place a "24/7 monitoring team" to address future incidents. While not specifically referring to the incident, Musk told the Microsoft event that xAI will practice transparency when mistakes are made.
This could be interpreted as a dig at archrival OpenAI, the creator of ChatGPT, which is Microsoft's main partner to build its in-house Copilot models. OpenAI, which was co-founded by Musk in 2015, is often criticized for keeping its technology's internal workings secret, as opposed to more open models like Meta's Llama or the technology from the Chinese company DeepSeek.
Oman Observer is now on the WhatsApp channel. Click here
