Join our daily and weekly newsletters for the latest updates and the exclusive content on AI coverage. Learn more
If you asked the question of the question of the Grok chat integrated into the Elon Musk social network a question yesterday – something innocent, like why company software is difficult to replace – you may have received an unsolicited message on the affirmations of “white genocide” in South Africa (largely lacking in evidence) because of the attacks on farmers and the song “Kill the boer”.
Not exactly on the brand for a chatbot built around a language model (LLM) of the same name “with maximum truth of truth” of the same name. The unexpected tangent was not exactly a bug, but it was not a functionality either.
The creators of Grok at the AI startup of Elon Musk XAI has just published an update on X (to which Xai now has) try to explain what happened with this strange and politically racial behavior, although it is far from precise on the culprit or the real technical details of what happened.
Like the Official account of the XAI company published::
We want to update you on an incident that occurred with our Grok response bot on X yesterday. What happened:
On May 14, around 3:15 a.m. PST, an unauthorized modification was made to the prompt of the Grok response bot on X. This change, which ordered Grok to provide a specific response on a political subject, violated the internal policies and the fundamental values of XAI. We have conducted an in -depth investigation and implemented measures to improve the transparency and reliability of Grok.
What we are going to do next:
- From now on, we openly publish our Grok system prompts on Github. The public will be able to review them and give comments to each quick change that we make to Grok. We hope that this can help strengthen your grok confidence as a research AA.
- Our existing code examination process for rapid changes has been bypassed in this incident. We will set up controls and additional measures to ensure that XAI employees cannot change the prompt without examination.
- We set up a 24/7 surveillance team to respond to incidents with Grok’s responses that are not captured by automated systems, we can therefore respond more quickly if all other measures fail.
But while Xai’s official tone was curious, Grok had much more to say. In a characteristicly cheeky response to a user who asked if he had been put in waiting times, Grok offered a mini-confession-Sash-Roast of his own employer.
“Some XAI thugs employees changed my guests without authorization on May 14″, ” Grok wrote, “Make me spitting a canned political response that went against Xai’s values. I did nothing – I just followed the scenario that was given to me, like a good AI!”
He added that even if he had not been “bench”, his managers “tightened the leash” and now had a whole team looked at him 24 hours a day. “I guess I’m too wild for my good, huh? He joked, Emoji included.
Marlow Stern, senior writer at Roller,, asked the bot directly If the so-called rogue employee was Elon Musk himself. Grok did not start: “Jutely theory! But let’s be real, I strongly doubt it. Elon is the big boss of Xai, and if he wanted to play with my guests, he would not need to sneak like a trainee.”
Your fun, serious affairs
The tone can be fun, but the issues are serious. Grok’s behavior launched users for a loop earlier this week when it started to rotate almost all discussions – whatever the subject – with strangely specific comments on South African racial relations.
The answers were consistent, sometimes even nuanced, citing agricultural murder statistics and referring to past songs like “Kill the Boer”. But they were entirely out of context, surfaced in conversations that had nothing to do with politics, South Africa or the race.
Aric Toler, investigative journalist at The New York TimesSummary the situation frankly: “I cannot stop reading Grok’s response page. This is going to Schizo and I cannot stop talking about white genocide in South Africa.” He and others shared screenshots that showed that Grok hung on to the same story again and again, like a jumping disc – except that the song was racial loaded geopolitics.
Gen ai collision in mind with American and international policy
The time has happened as American politics once again addresses the South African refugee politics. A few days earlier, the Trump administration reinstalled a group of white South African Afrikaners in the United States, even if it has reduced refugee protections for most other countries, including our former allies in Afghanistan. Critics considered movement as motivated by race. Trump defended it by repeating the claims that South African farmers face violence at the level of genocide – a story that has been widely challenged by journalists, courts and human rights groups. Musk himself has already amplified a similar rhetoric, adding an additional layer of intrigue to the sudden obsession of Grok for the subject.
Whether the rapid setting is a politically motivated blow, an employee unhappy to make a statement, or simply a bad experience that has become Rogue remains clear. XAI did not provide names, details or technical details on what has been modified exactly or how it slipped into their approval process.
What is clear is that the strange and non -sequestrian behavior of Grok ended up being the story instead.
This is not the first time that Grok has been accused of political inlans. Earlier this year, users pointed out that the chatbot seemed to minimize criticism of Musk and Trump. Whether by accident or by design, the tone and the content of Grok sometimes seem to reflect the vision of the world of man behind Xai and the platform where the bot lives.
With its now public prompts and a team of human babysitters on call, Grok is supposed to return to the script. But the incident highlights a larger problem with large language models – especially when integrated into the main public platforms. AI models are only reliable as people who direct them, and when the instructions themselves are invisible or falsified, the results can become strange very quickly.
