The rules of the game are well known: As soon as a generative artificial intelligence program goes online, internet users try to derail it. This is what happened when xAI, Elon Musk's artificial intelligence company, released version 2.0 of its Grok software on Tuesday, August 13. Accessible for "premium" subscribers to the social media platform X, the chatbot now not only generates text but also images. Internet users have since succeeded, with disconcerting ease, in getting it to produce violent images and other deepfakes.
There's a depiction of Musk in a high school, gun in hand, encircled by teenage corpses; Barack Obama preparing to consume cocaine; Donald Trump and Kamala Harris in swimsuits. Mainstream image generators generally prohibit such content. Dall-E (made by OpenAI), in particular, refuses to create representations of public figures or illegal activities.
In theory, these programs also prohibit copyright infringement. Yet Grok easily generates content featuring Mickey Mouse, Mario and SpongeBob SquarePants, for example – or even all three at once, as in this image where they share cannabis with Musk. While it also seems easy to use Grok to generate images of public figures in their underwear, the software does refuse to create pornographic content.
Grok is not the first AI system to enable the creation of this type of image. Nevertheless, the main commercial tools, with their strict restrictions, generally require the use of more or less sophisticated methods to get around them.
Grok's case is a special one: Musk has always vilified his competitors' generative artificial intelligence programs, which he claims are subject to political correctness and the "woke" agenda. He fiercely criticized Google in February, when its Gemini software portrayed Black people as German soldiers from the Second World War or the founding fathers of the US. "The woke mind virus is killing Western Civilization," he said, accusing Google of designing "racist, anti-civilizational programming."
While working on the launch of Grok, he confided to the Fox News channel that he wanted to develop a chatbot that sought "truth" above all else. When it was released in November, Grok, which could only generate text at the time, incorporated virtually none of the abuse protections that have become standard among its competitors.
Reacting to a message praising the "uncensored" aspect of Grok 2.0 and its supposed respect for "freedom of speech," Musk said on Wednesday: "Grok is the most fun AI in the world!," unleashing the comments of internet users who are fans of the multi-billionaire, who replied with images generated by Grok and featuring him, in heroic poses, alongside Trump, whose campaign he supports.
Regulators are unlikely to be amused by this new version of Grok, a few months ahead of the US presidential election, and at a time when the X social media network is the subject of a European investigation: The European Commission suspects it is failing in its obligations to moderate illegal content and misinformation.
Translation of an original article published in French on lemonde.fr; the publisher may only be liable for the French version.