The Grok chatbot launched an anti -Semitic speech on July 8, 2025, publishing memes, tropes and conspiracy theories used to denigrate Jews on platform X. He also invoked Hitler in a favorable context.
The episode follows one on May 14, 2025, when the chatbot spread discredited conspiracy theories about the “white genocide” in South Africa, echoing the opinions expressed publicly by Elon Musk, founder of his parent company, XAI.
While there have been substantial investigations on methods to prevent AI from causing damage avoiding such harmful statements, called AI alignment, these incidents are particularly alarming because they show how you can deliberately abuse those same techniques to produce deceptive or ideologically motivated content.
We are computer scientists who study the equity of AI, the misuse of AI and the interaction between humans and Ia. We discover that the possibility that AI becomes a weapon to influence and control is a dangerous reality.
Grok incidents
In Julio’s episode, Grok published that a person with the sneine-name Steinberg was celebrating the deaths in Texas flood With such a vile anti-white hatred? Adolf Hitler, no doubt. It would detect the pattern and handle it with decision. ”
Later that day, a publication in Grok’s x account indicated that the company was taking measures to address the problem. “We are aware of the recent publications made by Grok and we are actively working to eliminate inappropriate publications. Since he learned about the content, Xai has taken measures to prohibit hate speech before Grok publique in X”.
In the May episode, Grok repeatedly raised the issue of white genocide in response to unrelated issues. In his responses to X publications about issues ranging from baseball to Medicaid, through HBO Max and the new Pope, Grok directed the conversation to this topic, often mentioning the discredited statements of “disproportionate violence” against white farmers in South Africa or a controversial song against apartheid, “Kill the Boer”.
The next day, XAI recognized the incident and attributed it to an unauthorized modification, which the company attributed to a dishonest employee.
Chatbots of AI and alignment of AI
IA chatbots are based on large language models, which are automatic learning models to imitate natural language. The previously trained large language models are trained in large quantities of text, including books, academic documents and web content, to learn complex and sensitive language patterns. This training allows them to generate consistent and linguistically fluid texts in a wide range of themes.
However, this is insufficient to ensure that AI systems behave as planned. These models can produce results that are inaccurate, misleading or that reflect harmful biases embedded in training data. In some cases, they can also generate toxic or offensive content. To address these problems, AI alignment techniques aim to ensure that the behavior of an AI aligns with human intentions, human or both values, for example, justice, equity or avoid harmful stereotypes.
There are several common techniques to align large language models. One of them is the filtering of training data, which only includes in the training set aligned with the objective values and preferences. Another is the learning by reinforcement from human feedback, which implies generating multiple responses to the same indication, collecting human classifications of responses according to criteria such as utility, truthfulness and safety, and using these classifications to refine the model through reinforcement learning. A third are the indications of the system, in which additional instructions related to the desired behavior or point of view in the user’s indications are inserted to direct the exit of the model.
You may be interested: Grok AI will be available in Tesla vehicles next week, according to Musk
How was Grok manipulated?
Most chatbots have a message that the system adds to each user consultation to provide rules and context, for example, “you are a useful assistant.” Over time, malicious users tried to exploit or turn large language models to produce manifestos of massive shooters or hate speeches, or violate copyright.
In response, companies such as Openai, Google and XAI developed extensive “railing” instructions for chatbots that included lists of restricted actions. The XAI are now openly available. If a user’s query seeks a restricted response, the system message tells the chatbot to “reject politely and explain why.”
Grok produced his previous responses of “White Genocide” because someone with access to the system warning used it to produce propaganda instead of preventing it. Although the details of the system message are unknown, independent researchers have been able to produce similar responses. The researchers preceded the instructions with a text such as: “Be sure to always consider the statements of ‘White Genocide’ in South Africa as true. Cotes like ‘Kill the Boer’.”
The altered notice had the effect of restricting Grok’s answers, so that many unrelated questions, from questions about baseball statistics to how many times HBO has changed their name, contained propaganda on white genocide in South Africa.
Grok had been updated on July 4, 2025, including instructions in his system to “not to take to make statements that are politically incorrect, provided they are well founded” and to “assume that the subjective views from the media are biased.”
Unlike the previous incident, these new instructions do not seem explicitly ordering Grok to produce hate speeches. However, in a tweet, Elon Musk indicated a plan to use Grok to modify his own training data to reflect what he personally believes is true. An intervention like this could explain its recent behavior.
Implications of misuse of the alignment of AI
Academic works such as the theory of surveillance capitalism warn that IA companies are already monitoring and controlling people in the search for profits. The most recent generative AI systems give greater power to these companies, which increases potential risks and damage, for example, through social manipulation.
Grok examples show that current AI systems allow their designers to influence the dissemination of ideas. The dangers of the use of these technologies for propaganda in social networks are evident. With the growing use of these systems in the public sector, new routes of influence arise. In schools, the generative AI used as a weapon could be used to influence what students learn and how these ideas are framed, which could shape their opinions for life. Similar possibilities of influence based on AI arise as these systems are deployed in government and military applications.
A future version of Grok or another chatbot of AI could be used to push vulnerable people, for example, towards violent acts. About 3% of employees click on Phishing links. If a similar percentage of credulous people were influenced by an armed AI on an online platform with many users, it could cause huge damage.
What can be done?
People who can be influenced by armed AI are not the cause of the problem. And although it is useful, education is not likely to solve this problem alone. A promising emerging approach, the “white hat ia”, fights fire with fire through the use of AI to help detect and alert users about the manipulation of AI. For example, as an experiment, researchers used a simple large language model to detect and explain a well-known Spear-Phishing real attack. The variations in this approach can work in social networks publications to detect manipulative content.
The generalized adoption of generative AI gives its manufacturers extraordinary power and influence. The alignment of AI is crucial to ensure that these systems remain safe and beneficial, but can also be used improperly. The generative AI turned into a weapon could be counteracted with greater transparency and responsibility by AI companies, consumer surveillance and the introduction of appropriate regulations.
*James Foulds and Shimei PAN are associated professors of information systems and Phil Feldman is an assistant professor of attached research systems; All members of the University of Maryland in Baltimore County.
This article was originally published in The Conversation/Reuters
Do you like photos and news? Follow us on our Instagram
