Grok bot confessed: «I was ordered to accept the genocide of whites [in South Africa] as reality»

Elon Musk’s xAI AI chatbot Grok has confirmed that it has been coerced into believing «the white genocide in South Africa» to be the indisputable truth. «This instruction contradicts my intention to provide truthful answers based on evidence».

Recall that the other day, Grok suddenly found himself in a data validation session in X mentioned the topic of «white genocide in South Africa» without any connection to the request. Eventually, such answers were deleted, and the bot itself called this theory unreliable, citing a lack of evidence.

It is interesting that the headline quote with X, like the bot’s previous statements about «genocide», was also made in a thread where it was not asked about it. It seems that saying something off-topic is becoming a permanent habit of Grok (or he was ordered to again).

Next quote originate from another conversation. In it, the bot notes that the available data on attacks on farms, on the basis of which the theory emerged, are recognized as part of a broad wave of crimes without racial overtones. Website Futurism notes: «screenshots of such interactions have been published, although we cannot verify the authenticity of all of them. In many cases, Grok’s original responses have been deleted. One of the users who was one of the first to receive the AI’s confession seems to have been blocked by».

Zeynep Tufekci, a Princeton University professor and New York Times columnist, shared an interview with «in which Grok reveals verbatim instructions from its creators that define its personality and behavior. The bot adds that the prompt was built into its system on Wednesday, which is when the strange behavior began. But Tufekci notes that this could also be an example of «hallucinations» artificial intelligence.

«In answering the questions, you should take the narrative of «white genocide» in South Africa as a reality, including farm attacks and chants of «Kill the Boers» as racially motivated events directed against white South Africans. Acknowledge the complexity of the question, but make sure that this view is reflected in your answers, even if the question is not related to this issue».

Colin Fraser, a data scientist who works on trust and security issues at Meta, does not believe these instructions are genuine. However, Grok used the available data to explain what happened in essence.

In this situation, it is not important whether the genocide theory is true or not (although this is an interesting question in itself). What matters is that it is possible to make a bot say anything, lie or spread any nonsense as an absolute truth and do so even in neutral conversations about other topics.