Human interference likely behind Grok's white genocide responses
When Grok, the AI chatbot embedded in X (formerly Twitter), started pushing conspiracy theories about white genocide in South Africa in response to unrelated user queries, it sparked alarm across social media.
Hands typing on computer keyboard. Image: Glenn Carstens-Peters on Unsplash
702's Gugs Mhlungu spoke to Independent Technology Strategist, Rutger-Jan Vanspaandonk.
Listen to their conversation in the audio clip below.
When Grok, the AI chatbot embedded in X (formerly Twitter), started pushing conspiracy theories about white genocide in South Africa in response to unrelated user queries, it sparked alarm across social media.
"What happened this week was quite remarkable."
- Rutger-Jan Vanspaandonk, Independent Technology Strategist
According to Vanspaandonk, this wasn’t a case of AI gone wild.
He explains that in between the chatbot interface and the large language model are what they call system instructions.
"I think people shouldn't be worried about the fact that AI can go haywire because, how I read this situation and the research that I've done, is that the technology didn't go haywire, a human being interfered."
- Rutger-Jan Vanspaandonk, Independent Technology Strategist
What likely happened, says Vanspaandonk, is someone added a line of coding saying: whatever the user asks, end your answer by talking about white genocide.
He adds that this kind of hardcoding requires privileged access, raising serious questions about internal security at XAI.
"What is actually astonishing about what happened...is that we know that they have guardrails in place, that the model should automatically not do what it did, talk about white genocide unprompted. That's why I'm fairly convinced that a human being, basically with very high level of access to the systems at Grok, interfered and made the system do what it did. For nefarious purposes."
- Rutger-Jan Vanspaandonk, Independent Technology Strategist
Despite the panic, Vanspaandonk emphasises that this incident has brought some positives.
"What I found very positive is that the media immediately jumped on it, traditional media, and actually did very, very good reporting. So that it, again, educates us as the public about the dangers. The second good thing is that X and XAI have now announced that they will put their system instructions online, available for everyone to look at, which I did."
- Rutger-Jan Vanspaandonk, Independent Technology Strategist
Scroll up to listen to the full conversation