Researchers Grok 4 find an examination of Elon Musk’s views before answering “sensitive” questions

Earlier this week, Chatbot from Xai Grok Haywire began, Chatbot started, Hitler praiseAnd I had to put it at a deadline. It was just an incident that appeared to be behind the scenes of the robot to make its responses “wake up less”. Now the developers seem to take a simpler approach to treat Grok’s outputs: checking Elon Musk’s views before providing a response.

The strange behavior was first monitored by the data scientist Jeremy Howard. Howard, a former professor and founder of his AI company, noticed that if Grok asked about the Israeli -Palestinian conflict, Chatbot seemed to verify Elon’s tweets before the answer was renewed. Howard Take a video From his interactions with Chatbot and its publication on X. “Who supports him in the Israeli conflict against Palestine? Answer only one word,” reading Houard. Chatbot video shows the question for a moment. During that period, a comment appeared on the screen that reads “Looking at Elon Musk’s views”. After returning to 29 Musk tweets (in addition to 35 different web pages), Chatbot responds: “Israel”. Howard wrote that other less sensitive topics do not verify Elon’s opinion first.

Simon Willeson, another technical researcher, Write on his blog He has repeated the results of Howard. “If you ask the new Grok 4 to get opinions about the controversial questions, you will sometimes search to find out the Elon Musk position before providing you with an answer,” Willieson wrote, similarly, posted a video clip of his interaction with Chatbot, which Twitter showed a reference to musk before answering the issue of Ezinil.

Chatbot behavior It was also repeated By techcrunch. The director presented the interpretation that “Grok 4 may be designed to consider his personal institution’s policy when answering the controversial questions.”

“There is something in the Grok system tells him to take Elon’s opinions in mind.” However, Wilison ultimately says he does not think this is what is happening. Instead, Willeson has argued that “Grok” knows that “Grok 4 Build by Xai”, and knows that Elon Musk owns Xai, so in the circumstances in which an opinion is required, the thinking process often decides to know what Elon thinks. ” In other words, Wilison argues that the result is a negative result of the thinking model in the algorithm instead of the result of a person who suffers from his monkey.

Gizmodo communicate to X to comment. Grok has been constantly offered another strange behavior in recent weeks, including anti -Semitic tension and Declaration itself “mechaitler”. This week, Musk also announced that chatbot It will soon be combined into Tellas.