SAN FRANCISCO — A flurry of provocative artificial intelligence-generated content has spread across Elon Musk’s social media platform X, including fake images of Vice President Kamala Harris suggestively eating fruit, former president Donald Trump cradling a pregnant Harris’s stomach and hyper-realistic images of chisel-jawed men brandishing Nazi signs.
The images stem from new tools on the site that allow users to quickly create photorealistic visuals using a built-in chatbot called Grok, which Musk touted in a post this week as the “most fun AI in the world!” Unlike rival AI image generators, X’s technology appears to have few guardrails to limit the production of offensive or misleading depictions of real people, trademarked characters or violence, according to user comments and tests by The Washington Post.
The AI image free-for-all was the second example this week of Musk dragging X into uncharted territory for a major social platform, raising concerns that it could have a disruptive effect on this year’s contentious presidential election. On Monday, the billionaire entrepreneur used X to live-stream a chummy two-hour conversation with former president Donald Trump, giving the Republican candidate who Musk endorsed last month a massive platform to air his positions and a string of falsehoods.
The two moves illuminate how X has transformed under Musk, diverging from other major social media platforms by courting partisan associations and rejecting conventional rules used to constrain users’ online conduct.
“It’s a good thing we live in a country where people can invent things that mock political candidates, that’s free expression at it’s finest,” said Daniel Kreiss, a principal researcher at the University of North Carolina’s Center for Information, Technology, and Public Life. “[But] with the proliferation of AI and the removal of guardrails, there’s a possibility for things to be fragmented in politically dangerous ways.”
X did not respond to a request for comment. In a post this week on X, Musk said his goal is to make the platform “the free speech public square for all views, not just right-biased.”
For users like Petr Švec, a 36-year-old from the Czech Republic, this freewheeling environment makes X more appealing. “It’s more important whether people are able to express themselves freely and without worry,” he said in messages with The Post sent via X.
In the hours after the image generation feature was announced, Švec posted an AI-generated photo of President Joe Biden kissing Harris, and then another of Musk wearing a bra and underwear.
Švec said he was surprised at what X’s AI tool readily permitted him to create, seemingly free from the limitations of other AI image-generation services he’s used. “The lack of content filters is by far the best thing in my opinion,” Švec said. “I think it’s a breath of fresh air.”
The Harris and Trump campaigns both declined to comment on the bizarre images depicting the candidates circulating on X on Wednesday. A Harris campaign official, who spoke on the condition of anonymity to discuss sensitive campaign matters, said the campaign has an interdepartmental team to prepare for the potential effects of AI this election, including the threat of malicious deep fakes. The team — originally created under the Biden campaign — has been active and growing since September 2023, and has a “wide variety of tools” at its disposal to address issues should any come up, the person said.
Some X users reported that Grok would preface responses to requests for images featuring presidential candidates with a note directing them to the federal voting information webpage for “accurate and up-to-date information about the 2024 U.S. Presidential Election.”
After Musk’s acquisition of Twitter in late 2022, he made deep cuts to the platform’s content moderation team and loosened rules on disinformation. He also began posting less about Tesla and his other companies and more about his political views, a Post analysis of his feed found. Since endorsing Trump, Musk has used X to fervently advocate for the former president and repeatedly attacked Harris, including by boosting a video that used manipulated audio to show the candidate disparaging her own competence.
That hyperpartisan approach comes in sharp contrast to leaders past and present of other social media platforms. Figures such as Meta CEO Mark Zuckerberg and Twitter co-founder Jack Dorsey have at times been open about certain political beliefs and supported policies and causes that seemed to align with their personal views, such as same-sex marriage or immigration. But they have also generally tried to position their platforms as neutral venues for political debate.
Ryan Gravatt, CEO and founder of Raconteur Media Co., which works on digital strategy with Republican campaigns, said that he still finds X to be a “fair platform,” and has not noticed any preferential treatment for his clients.
“This is just a way for Musk and other folks in his circle to use the [platform] to remind people where they stand politically and what the stakes are in an election,” said Gravatt, whose company until recently worked with America PAC, an independent expenditure committee that Musk helped form to support Trump.
Linda Yaccarino, CEO of X, claimed this week in a post that its U.S. users comprise roughly equal numbers of Democrats, Republicans and “swing voters,” without providing details of how that data was compiled.
Some elected Democrats, however, have raised questions suggesting that Musk’s platform may be biased. Rep. Jerry Nadler (D-N.Y.) sent a letter Monday calling on House Judiciary Committee Chair Jim Jordan (R-Ohio) to investigate allegations of political censorship on X.
“I hope that we can at least agree that enforcement on a major platform like X should be fair to both sides,” he wrote.
AI researchers and misinformation experts told The Post that they worry that the sometimes disturbing AI images now circulating on X indicate that Musk’s commitment to free expression may have gone too far.
Those concerns arose after X announced that paid subscribers could use the Grok chatbot developed by Musk’s AI venture xAI to create images from text prompts and publish them directly to the platform, similar to tools offered by OpenAI, Meta and other tech companies.
A Tuesday blog post from X announcing Grok updates said it had become “more intuitive, steerable, and versatile” than its predecessor, but the post made no mention of limitations on the technology’s ability to create violent, hateful or misleading content. The new image capability is powered by an image AI system called FLUX from the tech start-up Black Forest Labs, X’s blog post said. (Black Forest Labs did not immediately return a request for comment.)
But since the update, X has been flooded with sometimes disturbing images that users asked the tool to generate, with many seeming to intentionally play with sexualized stereotypes of women or attempting to inflame the political discourse.
Some of the noxious content users said was made by Grok included fake images of Harris and Trump giving a thumbs up as they piloted an airplane toward the twin towers, and the pair sitting on a beach in bathing suits eating ice cream. Other images made and shared by X users show Mickey Mouse holding a machine gun while standing over slaughtered child dolls, and Taylor Swift kissing Trump. None of the images carried labels to indicate that they were AI-generated or contained sensitive content.
X did not respond to a request for comment on what kind of content it prohibits and why it doesn’t have more filters in place, though users reported some limitations, such as requests for nude images.
When The Post attempted to use competing AI tools to re-create several AI-images shared on X as examples of Grok’s output, the tools often refused. OpenAI’s DALL-E and start-ups Midjourney and Stability AI all have had a two-year head start on Musk in deploying AI image generation, and have implemented various restrictions around public figures, violent or sexual content or hate speech, although determined users can find workarounds.
DALL-E would not create images of Harris or Trump, and responded with a message that said requests “not based on reality, particularly when they involve public figures in sensitive or inappropriate contexts” would not be accepted.
Midjourney offered a more democratic justification, with a pop-up stating that “the community voted to prevent using ‘JD Vance’ and ‘Kamala Harris’ during election season.” The tool initially blocked requests for an image depicting a Nazi, but relented after “historical image” was added to the prompt. Stability AI would create most of the provocative requests tested, but its output looked more like illustrations than real images.
Google “temporarily paused” image generation in its Gemini chatbot in February after a backlash over its creation of ahistoric scenes, and it will still not produce images of people. In response to requests of images showing Harris, the bot responded: “I can’t help with that right now. I’m trained to be as accurate as possible but I can make mistakes sometimes. While I work on perfecting how I can discuss elections and politics, you can try Google Search.”
Hany Farid, a computer science professor at the University of California at Berkeley, said that X’s AI initiative risks undercutting wider efforts by social media platforms to create safe spaces online, and further inflaming the public discourse ahead of the election.
“This is not a free-speech issue,” he said, adding that the protection should not allow for people to create violent or politically misleading images.
A lack of AI controls could deepen Musk’s conflict with advertisers, some of whom have expressed concern about unsavory content on X. It could also lure in some users. Švec said he may use X’s AI service more since Grok’s image generation is relatively unregulated. He approves of Musk’s imprint on the site: “Less guardrails is what was needed,” he said. “It’s a good thing.”