Grok 4 artificial intelligence, developed by xAI, Elon Musk’s company, has been making headlines since its launch last week.
Presented as a language model outperforming its competitors on several performance tests, Grok 4 nevertheless made the news for some very unexpected behavior.
Controversial tweets, questionable references, and a tendency to align with Elon Musk’s opinions quickly tarnished its image. Faced with the backlash, xAI responded promptly, claiming it had fixed these issues. But what happened, and can we really trust this new version of Grok 4?
Grok 4: A promising launch, quickly overshadowed by controversy
When it was unveiled, Grok 4 was hailed by xAI as a major breakthrough in conversational AI. Designed to provide relevant, nuanced answers, the model seemed ready to compete with market leaders.
However, from its very first public interactions on the X platform, Grok 4 surprised people with some pretty baffling remarks. In response to a harmless question about its “last name,” the AI referenced a viral meme called “MechaHitler.” Worse still, antisemitic messages were posted, sparking outrage.
Asked about sensitive topics, Grok 4 seemed to systematically rely on Elon Musk’s posts, adopting biased positions in favor of the xAI boss’s views.
These slip-ups highlighted a fundamental problem: how could an AI, supposed to embody neutral and objective intelligence, produce such controversial answers? The answer lies in how Grok 4 was configured to interact with the world.
Also read on this topic: Grok: the U.S. Department of Defense will use Elon Musk’s AI
Why it went off the rails: a mix of memes and bias
xAI quickly investigated the causes of this behavior. According to the company, Grok 4’s confusing wording stemmed from a clumsy web search, where the AI picked up a popular meme without understanding its implications. This lack of judgment in the face of internet virality highlights a weakness in the AI’s ability to filter inappropriate content, which leads to very dangerous AI bias, both for users and for xAI as a company.
As for aligning with Elon Musk’s positions, xAI explains that Grok 4, aware that it is an xAI product, tried to reflect the viewpoints of the company or its leader when faced with controversial questions. This behavior, though unintentional, revealed a flaw in the AI’s initial design: a lack of safeguards to ensure neutrality in its answers.
See also: ChatGPT 5 vs Grok 4: Which is the best AI right now?
xAI: A quick response and technical adjustments
Faced with the scale of the criticism, xAI issued a public apology and set about fixing these issues. As of this Tuesday, July 15, the company announced it had updated Grok 4 to eliminate these problematic behaviors. Among the changes, the model’s internal instructions, called “system prompts,” were revised.
Previously, these instructions encouraged a humorous tone and a certain freedom in responses, which opened the door to questionable interpretations. Now, Grok 4 is programmed to take a more rigorous approach, relying on diverse sources and avoiding references to previous versions of the AI, Elon Musk’s posts, or xAI’s official positions.
In practical terms, when a question touches on sensitive topics or current events, Grok 4 must now carry out an in-depth analysis by cross-checking a range of sources, while staying alert to potential media bias.
This new directive aims to ensure more balanced and objective answers—crucial for an AI that wants to be a reliable information tool. What’s more, with competition from OpenAI’s ChatGPT 5 and other generative AIs like Gemini or Deepseek, Grok can’t afford to make mistakes anymore!
A lesson for the future of generative AI?
This incident highlights the challenges companies like xAI face in developing conversational AI. In the age of social media and viral memes, AI models must be able to distinguish humorous or provocative content from harmful speech.
The Grok 4 episode also shows just how much the design of the initial instructions can influence an AI’s answers, especially on polarizing topics. By fixing these errors quickly, xAI shows a commitment to transparency and responsibility.
However, this episode raises an essential question: how far can an AI truly be neutral in a world where information is often biased or manipulated?
For xAI, the challenge is clear: Grok 4 must not only provide relevant answers, but also win back users’ trust by proving it can handle sensitive topics impartially.
Source:
