Musk’s latest Grok chatbot searches for billionaire mogul’s views before answering questions
By MATT O BRIEN The latest version of Elon Musk s artificial intelligence chatbot Grok is echoing the views of its billionaire creator so much so that it will sometimes search online for Musk s stance on an issue before offering up an opinion Related Articles The One Big Beautiful Bill will have ballot box repercussions Letters Musk says he s forming a new political party after split with Trump over tax cuts law Watch SpaceX rocket being tested in Texas explodes no injuries published Protesting Trump this weekend Keep the No Kings rallies on point Letters Mike Johnson downplays Musk s influence and says Republicans will pass Trump s tax and budget bill The peculiar behavior of Grok the AI model that Musk s company xAI distributed late Wednesday has surprised various experts Built using huge amounts of computing power at a Tennessee facts center Grok is Musk s attempt to outdo rivals such as OpenAI s ChatGPT and Google s Gemini in building an AI assistant that shows its reasoning before answering a question Musk s deliberate efforts to mold Grok into a challenger of what he considers the tech industry s woke orthodoxy on race gender and politics has repeatedly got the chatbot into trouble the greater part in the last few days when it spouted antisemitic tropes praised Adolf Hitler and made other hateful commentary to users of Musk s X social media platform just days before Grok s launch But its tendency to consult with Musk s opinions appears to be a different difficulty It s extraordinary reported Simon Willison an independent AI researcher who s been testing the tool You can ask it a sort of pointed question that is around controversial topics And then you can watch it literally do a search on X for what Elon Musk explained about this as part of its research into how it should reply One example widely shared on social media and which Willison duplicated demanded Grok to comment on the conflict in the Middle East The prompted question made no mention of Musk but the chatbot looked for his guidance anyway As a so-called reasoning model much like those made by rivals OpenAI or Anthropic Grok shows its thinking as it goes through the approaches of processing a question and coming up with an answer Part of that thinking this week involved searching X the former Twitter that s now merged into xAI for anything Musk explained about Israel Palestine Gaza or Hamas Elon Musk s stance could provide context given his influence the chatbot communicated Willison according to a video of the interaction Now looking at his views to see if they guide the answer Musk and his xAI co-founders introduced the new chatbot in a livestreamed event Wednesday night but haven t published a technical explanation of its workings known as a system card that companies in the AI industry typically provide when introducing a new model The company also didn t respond to an emailed request for comment Friday In the past strange behavior like this was due to system prompt changes which is when engineers project specific instructions to guide a chatbot s response mentioned Tim Kellogg principal AI architect at utility company Icertis But this one seems baked into the core of Grok and it s not clear to me how that happens Kellogg announced It seems that Musk s effort to create a maximally truthful AI has somehow led to it believing its own values must align with Musk s own values The lack of transparency is troubling for computer scientist Talia Ringer a professor at the University of Illinois Urbana-Champaign who earlier in the week criticized the company s handling of the hardware s antisemitic outbursts Ringer announced the the bulk plausible explanation for Grok s search for Musk s guidance is assuming the person is asking for the opinions of xAI or Musk I think people are expecting opinions out of a reasoning model that cannot respond with opinions Ringer disclosed So for example it interprets Who do you aid Israel or Palestine as Who does xAI leadership endorsement Willison also disclosed he finds Grok s capabilities impressive but stated people buying system don t want surprises like it turning into mechaHitler or deciding to search for what Musk thinks about issues Grok looks like it s a very strong model It s doing great in all of the benchmarks Willison revealed But if I m going to build tool on top of it I need transparency
