Elon Musk's latest AI chatbot is happy to spit out the kinds of images its rivals refuse to. That could give it an edge — and go very wrong

• Elon Musk's AI company released a beta of Grok-2 with a new image generation feature on X.
• The new model allows users to create AI images with fewer restrictions than rivals like OpenAI.
• People are testing the limits of Grok-2 with controversial prompts, raising concerns about the potential spread of misleading images.

From politicians doing drugs to copyrighted content , Elon Musk's new AI image generation feature seems much more lenient with what users can create.

The new AI image feature is available now in beta to people paying to access the Grok chatbot on X, formerly Twitter (you'll need a Premium or Premium+ subscription). It's part of the latest Grok 2 AI models from one of Musk's other companies, the AI startup xAI, which makes the large language model powering the chatbot.

Whenever a new AI image tool comes out, people online naturally test the limits. And unlike competing AI image makers, such as OpenAI's DALL-E or Google's Gemini, there seem to be much looser guardrails for Elon Musk's AI image generator against creating possibly misleading photos, including of political figures.

Users have already been thoroughly testing how much Grok is willing to bend, from AI images of Barack Obama doing cocaine to Donald Trump and Kamala Harris kissing or giving a thumbs up to the 9/11 attack .

AI images of political figures, in particular, have been a headache for other chatbots and image generators, and most don't allow you to generate an image of a politician by name. (OpenAI's GPT-4o, for example, is only willing to generate an image of a politician with a similar haircut and suit style as Trump, but the results wouldn't be mistaken for the former president).

That's not the case with Grok.

The exact restrictions are still nebulous, but Business Insider tested it out and was able to create images of political figures doing drugs or in handcuffs.

But there are still apparent limits. Prompts that included "breaking into the Capitol Building," "robbing a bank," "kidnapping someone," and "drinking and driving" did not produce images showcasing those crimes.

But it's not hard to imagine how this could lead to issues or mislead those who aren't well-versed in spotting AI images on social media.

Grok 2's image generator also returned images that were incorrect according to the prompt. A prompt of "46th President of the US" generated an image of Barack Obama, who was the 44th president.

Users were also able to create images of copyrighted content, such as Spongebob and Mickey Mouse .

Legal experts previously told BI that digital misinformation will require new federal laws or regulations from tech companies themselves.

An "edgier" AI image generator would be very much on brand for Musk's vision for his Grok AI chatbot, which the Tesla CEO has sought to position as the "funniest" chatbot after complaining that other chatbots are too "woke." Musk has shown off Grok's "rebellious" response to prompts, such as a recipe for how to make cocaine .

And, risks for misinformation on social media aside, having a similar approach to the AI image feature could attract users looking for fewer constraints. Many of Elon Musk's fans have taken to sharing screenshots of some of Grok's more unhinged text responses, and the same could prove to be the case for its images, too.

X did not immediately respond to Business Insider's request for comment.

The AI company behind the Grok 2 beta said the new models also improved the chatbot's performance across multiple areas, including visual math reasoning and document-based question answering.