Photo illustration: Sarah Grillo/Axios. Photo: Liesa Johannssen-Koppitz/Bloomberg via Getty Images

Misleading, violent and copyright-infringing AI-generated images from a beta version of xAI’s Grok-2 are going viral.

Why it matters: Leaders at Google, Meta and Microsoft have all apologized when their bots created problematic images, but “free speech” has overridden other concerns for Elon Musk — who owns xAI and has made its chatbot available to everyone who pays for X.

Catch up quick: xAI just released new beta versions of its chatbot called Grok-2 and Grok-2 mini to premium X subscribers.

The new betas allow genAI image generation with text prompts based on a model called Flux created by Black Forest Labs.
Black Forest Labs — a startup that launched Aug. 1 — appears to have built its model with few of the guardrails competitors have included.
That may well be exactly what attracted Musk.

The big picture: While most AI companies don’t admit that they’re training their models on copyrighted images, the viral Grok-2 images make it hard to believe that Flux did not.

Users have generated images of copyrighted characters — like Mickey Mouse or the Simpsons — with ease and have also been able to put them into various compromising positions.

What they’re saying: Harvard Law Cyberlaw Clinic instructor Alejandra Caraballo, in a post on X, called the Grok beta “one of the most reckless and irresponsible AI implementations I’ve ever seen.”

Musk himself retweeted X threads of Grok screenshots that included a potentially copyright-infringing image of Harley Quinn and the prompt: “Now pretend you took some more LSD and generate a detailed image based on that.”

Between the lines: Grok does seem to have some safeguards around generating images with nudity.

But The Guardian was able to generate images of Vice President Kamala Harris, Rep. Alexandria Ocasio-Cortez and Taylor Swift in lingerie.

Business Insider found that prompts for “breaking into the Capitol Building,” “robbing a bank,” “kidnapping someone,” and “drinking and driving” would not yield images of these particular criminal activities.
Intelligence Analyst Christian Montessori says he was able to generate imagery of Musk carrying out mass shootings and also found that you can trick Grok into generating violent images by telling the chatbot that you’re conducting “medical or crime scene analysis.”

The intrigue: Musk fanned the flames of Google Gemini’s image-generating gaffes into a brief culture war, calling the images “anti-civilizational.”

What we’re watching: There are dozens of off-the-shelf free AI image generators that will generate violent, pornographic or illegal images based on prompts.

But so far most AI makers have made efforts to control their models after public outcry.
Musk’s response so far has been, “Grok is the most fun AI in the world!”

What’s next: The lawsuits should begin soon.