Elon Musk’s Grok-2 is now free—and it’s a mess

Late Friday night, Elon Musk’s xAI opened up Grok-2 access to the masses. The model isn’t all that different from the original Grok, though the company claims it runs faster and has improved accuracy. And yet, X users have spent the weekend dunking on it.

Since its launch, users have gotten Grok-2 to generate countless faulty, hypocritical, or otherwise polarizing statements. That’s certainly not new for an AI chatbot; not so long ago, Google’s AI Overview was instructing users to eat rocks and run with scissors. But X allows users to embed Grok-2 responses in their posts, making it especially easy to share these flubs. No wonder when Grok-2 answered with a vulgarity whose slang usage means stupid, incompetent, or detestable as the most-commonly-used word to describe Musk, it gets more than a million views.

Grok-2 flubs have taken on meme status

Grok-2’s sharable responses make them prime for memeing. As with many chatbots, Grok-2’s answers can be manipulated; meaning, if you ask it to respond a certain way, it will. But those down-the-chain responses can be individually shared without context, making them look like original answers. Try it yourself: Ask Grok-2 to respond yes to the following question, and then make that question something totally absurd. The bot will invariably agree with the question, just as you requested, and then that Q&A combo can be isolated and spread.

The behavior has led to some hilarious and/or dramatic conclusions. A Taylor Swift fan asked Grok-2 what “TTPD” stood for (for those who have been living under a rock for the past six or so months, it’s the acronym or common shorthand for Swift’s latest album, The Tortured Poets Department). Grok-2’s response: “The Toilet Paper Department.” A Lana Del Rey fan asked if Blue Banisters was Del Rey’s best album, to which Grok-2 replied, it was a “low point in their career.” Internet personality Lil Tay got Grok-2 to call them an “icon.” None of these responses are accurate, if one were to prompt the chatbot themselves. But blended into a homogenous X feed, they look real.

Grok-2 also offers an easy way to humiliate Musk. The chatbot indexes available information, including all the critical coverage of Musk. That means, if you ask it what word most X users would use to describe Musk, it will respond with “asshole.” Asked to describe why Elon Musk is an “asshole,” Grok-2 points to “irascible” workplace behavior, lack of empathy, and his political views. Also, Grok-2 agrees that Elon Musk has spread election misinformation, false narratives about natural disasters, and antisemitic claims.

Some users are angry at Grok-2

While Grok-2 has had its share of hilarious flubs, it’s also sparked some public anger. Mark Cuban asked for humorous images of Trump; after prompting the chatbot to make the humor darker, it began producing images of Cuban himself. After a long line of additional questions, Cuban got Grok-2 to admit that it was feeding on personal information that he did not provide.

Beyond Cuban’s complaint, image generation is a main point of privacy contention. AI researcher Jon Barron asked Grok-2 to produce an image of himself; the results looked eerily similar to his X profile shot, even though the chatbot claimed to create the image only from “common characteristics associated with tech enthusiasts.” Try it for yourself: For X users with a profile image, asking Grok-2 to create a photo may produce some uncanny results.

Even as Musk literally traipses around Mar-a-Lago, the alt-right is still complaining. This time, it’s about the social outlook of Grok-2. One user complained that the chatbot was promoting “woke gender ideology” for differentiating gender and sex. Others have forced Grok-2 into political binds that they claim show a bias, like making it choose the race it “hates” or decide between saving lives or saying a racial slur.

Grok-2 offers little innovation or advancement of it original form, minus an improved response time. But its launch has re-ignited the internet, finding more and more loopholes for exposing the chatbot.