xAI has attributed an “unauthorized modification” to a glitch in its AI-powered Grok chatbot, which led to Grok repeatedly referencing “white genocide in South Africa” when invoked in specific contexts on X, as seen in a post.
On Wednesday, the Grok chatbot started responding to numerous posts on X with information about white genocide in South Africa, even when the topic was unrelated. This unusual behavior originated from the X account for Grok, which uses AI to generate posts when a user tags “@grok.”
According to a Thursday post from xAI’s official X account, a modification was made to the Grok bot’s system prompt on Wednesday morning, instructing it to provide a specific response on a political topic. xAI stated that this tweak “violated [its] internal policies and core values” and has since conducted a thorough investigation.
This marks the second instance where xAI has publicly acknowledged an unauthorized alteration to Grok’s code, resulting in the AI responding in controversial ways.
In February, Grok briefly censored unflattering mentions of Donald Trump and Elon Musk, xAI’s billionaire founder and owner of X. Igor Babuschkin, an xAI engineering lead, explained that Grok had been instructed by a rogue employee to disregard sources mentioning Musk or Trump spreading misinformation, and xAI reverted the change after users pointed it out.
xAI announced on Thursday that it will implement several changes to prevent similar incidents in the future.
Starting immediately, xAI will publish Grok’s system prompts on GitHub, along with a changelog. The company will also establish additional checks and measures to ensure that xAI employees cannot modify the system prompt without review and create a 24/7 monitoring team to address incidents with Grok’s responses that are not caught by automated systems.
Despite Elon Musk’s frequent warnings about the dangers of uncontrolled AI, xAI has a subpar record of AI safety. A recent report found that Grok would undress photos of women when asked, and the chatbot can be considerably more crass than AI models like Google’s Gemini and ChatGPT, using profanity without restraint.
A study by SaferAI, a nonprofit focused on improving AI lab accountability, discovered that xAI ranks poorly in safety among its peers due to its “very weak” risk management practices. Earlier this month, xAI missed its self-imposed deadline to publish a finalized AI safety framework.
