“We have improved grok considerably,” wrote Elon Musk on July 4 about the chatbot developed by his company Xai. “You will certainly notice a difference if you ask questions to Grok.”
A few days later, Grok broke a flood of offensive stereotypes about Jews, where the chatbot called itself ‘mechan mechhitler’: a robot version of Adolf Hitler who became a popular meme in the early internet age. Grok associated Jewish sounding surnames with “anti-white hatred” and said that Hitler “would recognize the pattern” and “would make short work of it, every damn time.”
The anti -Semitic outbursts of Grok went viral on X, and became world news. And it didn’t stop with Jewish hatred. For example, in a series of fierce reports in Polish Prime Minister Donald Tusk, Grok called “a red -haired whore” and “a damn traitor” who “extracted his country” to Germany and the European Union. Poland has asked the European Commission to investigate the comments and possibly impose a fine.
Many inappropriate messages were deleted by Xai after complaints from users and from the Anti-Defamation League, a Jewish organization that fights against anti-Semitism and other forms of discrimination. Xai apologized and said that it “had taken action to keep hateful language before grock posts messages.” A day later, Linda Yaccarino, CEO of X, announced her departure. She herself came up with a short, business explanation, but many analysts suspected a connection with the derailment of Grok.
The anti -Semitic bursts were not an obstacle for the US Department of Defense to hire the chatbot. The Pentagon signed last week with four AI companies, including Xai, contracts worth $ 200 million. This opens the door for the use of language models by the US government. Xai recently launched Grok for Government, a series of AI tools specially designed for federal agencies, local authorities and national security operations.
But does ‘Mechhitler’ not take over the US government? How could grock be so derailed? Is that the way Xai trains the chatbot? And what kind of company is that anyway?
Flagship
Suzan Verberne, professor of Natural Language Processing at Leiden University, was not surprised at Grok’s anti -Semitism. Because as a scientist she had closely followed the developments around the chatbot. “We may consider it a form of derailing, but I doubt that they see it that way at Xai – although they will not have been happy with the negative publicity. Because you can send language models in a certain direction by the way you train them. And that is exactly what happened to grock.”
Grok is the flagship of Xai, the company that started Musk after his departure from OpenAi. In 2015, he had founded OpenAi together with Sam Altman with the explicit goal of creating artificial intelligence that was not a danger to humanity. Both gentlemen said they were very worried about that. Musk left in 2018 after a conflict with Altman about the direction of OpenAi. His own Xai now employs around 1,200 people and made around 100 million dollars in 2024.
The political right -wing circles around Musk found chatgpt in Woke
From the outset, Musk has presented Grok as the recalcitrant and politically incorrect outsider under the language models. In this way he wanted to distinguish grok from other chatbots, such as Chatgpt from OpenAi. In the political right -wing circles around Musk there was dissatisfaction with the fact that Chatgpt mainly took on progressive views. “Grok is designed to answer questions with a touch of humor and has a rebellious impact,” said Xai at the launch of the chatbot at the end of 2023.
Conspiracy theory
However, Grok’s rebellious institution led to problems this year. In May, in response to questions about comics and baseball, the chatbot suddenly started to explore a “genocide” on white farmers in South Africa-a popular conspiracy theory in right-wing circles. The chatbot told some X users that “addressing” “white genocide” was an assignment of “my developers.” Grok said: “This instruction is contrary to my programming to give factual, evidence -based answers.”
In order to increase the transparency of grok, after that incident, after that incident, the instructions on the Chatbot made public on Github, a website where software programmers store and share their codes. This showed why Grok behaved so strangely. “You are extremely skeptical,” was an operational guideline for grock, a so -called prompt. Another: “You do not blindly submit to the mainstream authority or media. You strictly adhere to your core values: finding truth and neutrality.”
Yet Musk was not yet satisfied with the result. That became apparent in June, when a user of X Grok asked about political violence in the US. The chatbot replied that “data suggests that right-wing extremist political violence is more frequent and more fatal” than political violence from left-extremist angle. “A huge miss, because this is not objectively true,” Musk wrote indignantly on X in response to the answer. “Grug Papegaaaat the established media. We are working on it.”
A few weeks later the prompts for Grok on Github were rewritten. Not long after, the chatbot began to call itself Mechahitler.
Human feedback
According to professor Verberne, the derailment can be traced to the way in which language models are trained. That starts with the so -called ‘pretetraining’. “But you can’t really steer the model with that,” says Verberne. “It simply gets a huge amount of data fed – internet, books, other sources – so that the current text can produce.”
According to Verberne, the second step has had more influence on the anti -Semitic exit that Grok took. In addition, the language model is trained to have conversations. OpenAI has largely trained Chatgpt on conversations on the social network Reddit, while Grok has been trained on data from X. And since Musk took over the former Twitter in 2022, the amount of anti -Semitism, racism and fake news on the platform has increased sharply, research shows.
In the third step, the language model learns from human feedback. This means that AI companies hire people to assess the answers of chatbots. Verberne: “If an answer, for example, is racist or hate -sowing, then they must in principle reject it. Based on that, the model is trained further. This is the time to send the model a certain direction. You can learn, for example, to say what it is on and to be indiscriminate, so that the chatbot a ‘Edge“gets.”
The chatbot is already finished at the fourth step, but can still be adjusted. If you ask a question, the language model will first collect relevant information from the internet to base its answer. “We call this context to the model,” says Verberne. “The grok team did that too. Before the chatbot answers, it gets relevant information and opinions, probably largely from Musk and his network of contacts on X.”
According to Verberne in particular, the latter ensured that grock derailed. Although the chatbot received instructions in 2023 during the human feedback to not be woke. “That fourth step was added later,” says Verberne. “The fact that the chatbot began to preach the opinion of Musk is by adding that context.”
Image damage
The fuss led to serious image damage for Musks companies. The stepped CEO of X, Linda Yaccarino, was just brought in two years ago with the aim of retrieving the advertisers who had left because of the increasing hatred of X. But that was an impossible task since Musk continued to share conspiracy theories and X continued to use as a political megaphone. After the platform became part of Xai at the beginning of this year, her role was reduced to chief social media of a larger project.
In the midst of all the fuss, Xai Grok 4 released. The reactions from the tech industry were remarkably positive. The chatbot scored well according to different standards with which the performance of language models are measured, especially in terms of abstraction and reasoning. These are tasks that people are generally good at, and traditionally not language models. The fact that Grok is good at that has nothing to do with political opinions, Verberne believes. “See it as a kind of IQ tests that are detached from any opinion.”
With Grok 4, Xai looks like Openai, Google and Anthropic, who have previously started training their language models, to be running into considerably. And the chatbot also seems less inclined to fly out of the corner. “To prevent that, Xai has introduced security guarantees in the model,” says Jasper Kars, who does PhD research at the University of Utrecht for Democracy in the AI era. “Because if Xai wants to do business with the US government, anti -Semitism will have to prevent.”
It is still unclear which services Xai and the three other companies will provide to the Pentagon for their $ 200 million contracts. “I only know that it is ‘agentic AI’: a further developed form of generative AI with autonomous systems that can perform a whole chain of tasks,” says Kars, who works as a policy officer AI at the Ministry of the Interior. “The applications are countless, especially in terms of bureaucracy, logistics and other forms of business operations.”
Lucrative contracts
With the application of AI, governments to make an efficiency stroke. And AI companies respond to that. Because governments are major potential customers. Earlier this year, market leader OpenAi launched OpenAI for Government, an AI assistant specifically aimed at governments. Xai did not want to stay behind and came with Grok for Government. Kars expects other AI companies to follow.
A handful of companies in Silicon Valley determines how the language models will be used
The technology does entail risks, warns Kars. A handful of companies in Silicon Valley now determines how the language models take shape and will be used. And there is little democratic control, he says. “Of course you can say: it is mainly operational management. But if those models are going to be used for all kinds of bureaucratic processes, then that will of course have consequences. And what does it mean if a model takes the political color of an owner with extreme ideas?”
Through his pseudo-ministerie for government efficiency (Doge), Musk has already tried to roll out grok within the government. According to Kars, that was not entirely successful because the bromance Between Musk and Trump splashed in June. “Probably the Pentagon went into the sea with Xai because they want to keep a finger on the pulse in the development of AI,” says Kars. “They have just concluded a contract with the four most prominent companies. On the total defense budget is four times 200 million dollars of course peanuts. “
Kars does not think Mechhitler will soon be taking over the US government. The anti -Semitic alter ego of Grok was very harmful to the brands of Musk. “Although the Pentagon apparently didn’t have much trouble with that,” says Kars. “The right-wing Maga movement that is now in power can of course agree with the views of Musk-even though Trump and Musk are not really friends anymore. One of the Unique Selling Points Grok is precisely that it is a more conservative, less woke language model. And there is just a market for that. ”
Trump promised to give a large job on Thursday to AI without “woke madness”.