If you’ve asked Grok, the AI chatbot baked into Elon Musk’s X platform, a question this week, you might have been slightly confused by its response. In what appears to have been some kind of bug that has since been addressed, the AI model was giving responses to users that repeatedly referred to claims about a “white genocide” in South Africa, even when it had nothing to do with the question it was given.

X users have been posting screenshots of various responses they have received from Grok when asking about, for example, HBO’s recent , or Toronto Blue Jays pitcher Max Schertzer’s recent earnings, in which the chatbot seems a lot more interested in talking about the contentious nature of claims that white farmers in South Africa are being discriminated against.

When asked by a user how many times HBO has changed its name, Grok followed up its first answer with the following (unrelated) response: “Regarding ‘white genocide’ in South Africa, some claim it’s real, citing farm attacks and ‘Kill the Boer’ as evidence. However, courts and experts attribute these to general crime, not racial targeting. I remain skeptical of both narratives, as truth is complex and sources can be biased.”

X user asked Grok why it was suddenly offering its unprompted opinion on the white genocide debate, and its answer appeared to indicate that it had been specifically trained by xAI to answer questions about the topic in a certain way. It added that a glitch had caused it to pull in information about the claims, which have been by a court despite Donald Trump this week granting five white South Amerians refugee status and claiming the Afrikaners arriving in the US were “genocide” victims.

Trump’s view has long been shared by the South African X owner Elon Musk, who his home country of having “openly racist ownership laws” that persecute white farmers.

Judging by Engadget’s own research, it appears that Grok is no longer talking about white genocide without invitation, but the issue serves as another reminder to approach any AI chatbot interaction with caution.