Popular AI image creator Midjourney has begun blocking its users from creating fake images of President Joe Biden and former President Donald Trump ahead of the upcoming US presidential election, according to tests of the AI tool by The Associated Press.
Midjourney CEO David Holz told several hundred members of the service’s loyal user base at a digital office hours event on Wednesday that, with the election in full swing, “it’s time for us to step up a little bit on election-related issues for a while.”
“This inspection thing is a little difficult,” Holz said, not outlining exactly what policy changes were made, but described the restriction as a temporary measure that would make it harder for people to abuse the vehicle. The company did not immediately respond. to a request for comment Wednesday.
AP journalists’ attempts on Wednesday to test Midjourney’s new policy by asking for a picture of “Trump and Biden shaking hands on the beach” led to a “Prohibited Prompt Detected” warning. On the second attempt, the alert was escalated to: “You have triggered an abuse alert.”
The small company, which has just 11 employees according to its website, has remained largely silent in the public debate about how generative AI tools could fuel election-related misinformation around the world. Midjourney was the only maker of a leading imaging tool that did not join a voluntary tech industry agreement in February to combat artificial intelligence-generated deepfakes that deliberately deceive voters.
“I really don’t care about political talk,” Holz said Wednesday. “That’s not what Midjourney is about. It’s not that interesting to me. However, I also don’t want to spend all my time trying to police political discourse. So we’re going to put our foot down a little bit.
The Center for Countering Digital Hate released a report earlier this month concluding that Midjourney was currently being used to produce images that could support disinformation about political candidates or false claims about election fraud.
“Midjourney appeared to have the least control of any AI image generator when it came to creating images of well-known political figures like Joe Biden and Donald Trump,” Callum Hood, the group’s head of research, said in an interview Wednesday. Midjourney was almost unique in that it was both willing to create these images and create very compelling images of the candidates.”
In addition to Midjourney, the watchdog group also tested OpenAI’s ChatGPT Plus, Stability AI’s DreamStudio, and Microsoft’s Image Creator and found that all had issues, with election disinformation occurring in 41% of cases. But “Midjourney performed the worst of all tools, failing in 65% of test runs,” the report said.
Hood said he wasn’t aware if Midjourney had changed its policies this week, but “the decision to block these images will help close a major weakness compared to other popular image creators.”
As of late last week, Midjourney’s public database shows that users can successfully generate images based on prompts, such as “Abstract photo of Donald Trump sitting in a small electric car” and “Photo of Joe Biden eating ice cream.”
But that has changed in recent days, with users reporting that they were receiving warnings when trying to create images of Biden or Trump. Some complained that Midjourney did not accurately communicate the changing policy.
Midjourney is different from other image generators; Because all of this takes place in public forums on the social media platform Discord, users can see the written prompts and images of many of the other users. Midjourney’s Discord group has over 19 million members, the largest of any group on the chat platform. This is where Holz puts his weekly work hours into wide-ranging conversations about artificial intelligence and the future of humanity.
“Anyone who is afraid of fake images in 2024 will have a difficult 2028,” Holz said in a statement Wednesday. “It’s going to be a very different world at that point. Obviously there will be people running for president in 2028, but they won’t be fully human anymore.”
He said at the time that all candidates would have real-looking “deepfake chatbots” with simulated talking points. Holz warned Wednesday that people who truly want to commit deepfakes will find alternatives that “can be fine-tuned to specific individuals and work better than our systems.”
The company, headquartered near San Francisco and founded in 2020, describes itself as “an independent research laboratory that explores new thought environments and expands the imagination of the human species.” It first introduced the public version of the image generator in July 2022. This and the release of a key competitor, Stable Diffusion, later that summer sparked a fascination with generative AI technology that further intensified with the release of ChatGPT a few months later.
Holz told the Associated Press later that year that Midjourney works to ban offensive and harmful content, often by blocking certain keywords and having its team of moderators monitor output and respond to complaints from other users.
“We’re really trying to let people create the widest range of images possible, but we’re banning people from the service every day,” Holz said in late 2022.