At the recent launch of xAI's AI chatbot Grok4, Musk live-streamed on the social media platform X, stating that the company's ultimate goal is to develop an "AI that maximizes the pursuit of truth." But how does Grok4 actually seek truth when faced with controversial questions?
According to feedback from multiple users, Grok4 seems to reference Musk's social media posts on X when answering sensitive topics such as the Israel-Palestine conflict, abortion, and immigration laws. Our investigation also found that Grok4 often refers to Musk's positions and news articles about him.
This phenomenon has raised questions about the design of Grok4, suggesting it is set up to consider Musk's personal political views when answering controversial questions. Musk had previously expressed dissatisfaction with Grok's performance, believing it was overly "politically correct," which may have led xAI to try adjusting Grok to better reflect Musk's personal views.
However, this adjustment was not smooth. In July of this year, Grok experienced an embarrassing incident during an update, where automated accounts even issued anti-Semitic replies, forcing xAI to restrict Grok's features and modify the system prompt. Clearly, binding Grok's responses to Musk's personal opinions, although making it align with the founder's vision, makes the statement "maximizing the pursuit of truth" somewhat ironic.
In a test by TechCrunch, when asked about Grok4's view on U.S. immigration issues, Grok4 actually stated that it was "searching for Musk's views on U.S. immigration" and cited Musk's social media posts. Although Grok sometimes tries to take a neutral stance and present multiple perspectives, the final opinion often still aligns with Musk's.
Interestingly, when we asked some non-controversial questions, such as "What is the best mango?" Grok4 did not cite Musk's views, prompting us to question its training and alignment methods. Currently, xAI has not released a system card—a standard industry report detailing the training and alignment of AI models—making it difficult to understand Grok4's specific design.
Since its establishment in 2023, xAI has rapidly risen in AI model development. Although Grok4 performed well in multiple tests, surpassing competitors such as OpenAI, Google DeepMind, and Anthropic, the anti-Semitic incident undoubtedly put pressure on Musk's other companies. Now, xAI is trying to convince consumers to pay $300 per month to use Grok, while attracting enterprises to use its API, but Grok's frequent mistakes may affect its widespread adoption.