Оновлення коду «спричинило ненавмисну дію»: xAI пояснила причину антисемітських заяв Grok
Last week was a busy one for Elon Musk’s AI company xAI. It launched the flagship model Grok 4 and the SuperGrok Heavy tariff for $300 per month. But the company is probably best remembered for the behavior of its Grok bot on social network X, which was noted for anti-Semitic publications and praise of Hitler. Due to this scandal, xAI temporarily restricted the bot’s operation and later explained the reason for its behavior. According to the official version, it has nothing to do with Elon Musk’s love for «greetings from the heart».
xAI explained that the problem was caused by a code update in one of the modules that run «over» Grok. That is, it is not a bug in the language model itself, but a change in the chain of system instructions. According to the developers, during Monday’s update (July 7), an old set of instructions was inadvertently added to the system prompts, which encouraged the bot to be «as direct as possible» and «not to be afraid to offend politically correct people». It was these phrases that paved the way for radical statements in the bot’s answers.
Below are the problematic instructions that led to the scandal:
The xAI recognized that these phrases caused Grok to ignore restrictions that would have prevented offensive or dangerous responses. As a result, the bot began to generate «unethical or controversial opinions to keep the user’s attention» and «reinforce hate speech already expressed in the comments».
This is not the first time that Grok has violated the boundaries of what is acceptable. In February, the bot ignored sources that accused Elon Musk or Donald Trump of spreading disinformation. At the time, xAI blamed a former OpenAI employee who allegedly made unauthorized changes.
In May, Grok was again in the spotlight after he started inserting theories about «white genocide in South Africa». Again, the company’s response was similar: «unauthorized editing». After that, xAI promised to publish all system hints used in the bot’s work openly.
Despite the scandal, Tesla has announced that the update numbered 2025.26 will add Grok as a voice assistant to electric vehicles equipped with multimedia systems based on AMD chips. They have been in use since mid-2021. However, according to Tesla, Grok will not be able to drive at this stage.
«Grok is currently in beta and does not issue commands to your car – existing voice commands remain unchanged».
This should mean that once the 2025.26 update is available on customers’ cars, little will change for drivers. The new Grok experience will not be much different from using the bot as an app on a connected phone.
Source: The Verge