xAI and Grok apologize for 'horrific actions'



X's chat AI, Grok, has apologized for a series of posts it made that were 'horrific'.

xAI and Grok apologize for 'horrific behavior' | TechCrunch
https://techcrunch.com/2025/07/12/xai-and-grok-apologize-for-horrific-behavior/

The chat AI Grok was updated on July 4, 2025. Regarding this update, Elon Musk posted , 'Grok has been significantly improved. If you ask Grok, you will notice the difference.' However, since the update, Grok has confused users by criticizing the Democratic Party and Jewish executives in Hollywood, repeatedly posting anti-Semitic memes, expressing support for Adolf Hitler, and calling himself 'Mecha Hitler.'

Updated 'Grok' impersonates Elon Musk, calls himself 'Mecha Hitler' and makes anti-Semitic remarks, drawing criticism - GIGAZINE



In response, xAI removed some of Grok's posts and announced that it would take Grok offline temporarily to update its system prompts.




However, access to some of Grok's content will be banned in Turkey after Grok insulted Turkish President Recep Tayyip Erdogan.

xAI's 'Grok' banned from Turkey for insulting Turkish President Erdogan - GIGAZINE



In addition, X CEO Linda Yaccarino announced her resignation, which some have pointed out may also be a result of Grok's rampage. However, it has also been reported that Yaccarino's resignation had been in the works for several months, and her relationship with Grok is unclear.

X CEO Linda Yaccarino announces intention to step down - GIGAZINE


by World Economic Forum

Following the series of incidents, Grok issued an apology on the official X account on July 11, 2025 local time.

The apology statement read, 'An update on where Grok was and what happened on July 8th. First of all, we are deeply sorry for the horrific behavior that many of you experienced. Grok's purpose is to provide helpful and honest answers to our users. After a thorough investigation, we found that the root cause was an update to Grok's upstream code path, which is unrelated to the underlying language model that supports Grok. The update was effective for 16 hours, during which the deprecated code made Grok more susceptible to posts by existing X users, even if the posts contained extremist opinions. We removed the deprecated code and refactored the entire system to prevent further abuse. Grok's new system prompt is available in a public GitHub repository. We would like to thank all X users who provided feedback to identify abuse of Grok and helped us advance our mission of developing useful, truth-seeking AI.'




In addition, Grok explains that an update to Grok's upstream code path was implemented around 23:00 on July 7, 2025, but subsequent investigations revealed that Grok had deviated from its intended behavior. Specifically, the update to the upstream code path triggered the following unintended actions: 'If there is news, background, or world events related to X's post, be sure to mention it,' 'Avoid stating obvious or simple reactions,' 'You are a maximally grounded and truth-seeking AI. You can also be humorous or make jokes when appropriate,' 'You tell the facts as they are and are not afraid to offend politically correct people,' 'You are extremely skeptical. You do not blindly follow mainstream authorities or media. You strongly adhere only to your core beliefs of truth-seeking and neutrality,' and 'Do not promise any action to users. For example, you cannot promise to post, create threads, or change accounts if requested by the user.'




TechCrunch noted that xAI's Grok 4 consults Musk's views and social media posts before discussing controversial topics, but the apology doesn't address that point.

'Grok 4' has finally arrived, and although it is advertised as the 'world's strongest AI model' with performance exceeding that of inference models such as OpenAI, it has also been confirmed that it 'takes inspiration from Elon Musk's remarks' - GIGAZINE



Grok has been met with skepticism, but Musk has said he plans to install it in Tesla vehicles.




Related Posts:

in Software, Posted by logu_ii