Content warning: This article contains numerous examples of bigoted rhetoric.
Users on the far-right message board site 4chan have used Bing’s AI image generator to create numerous images promoting Nazi imagery and “propaganda” seemingly designed to inflame opinion surrounding the conflict in Israel and Gaza, despite promises from the companies involved that they were working to address the issue.
Microsoft, which owns Bing, says it prohibits harmful content, and OpenAI, which created the technology powering Bing’s image generator, prohibits “hateful, harassing, or violent content.” But Media Matters found hundreds of cases of users sharing a link to Bing’s image generator, with over 100,000 combined replies — many of them including an apparent AI image.
On October 3, Microsoft launched DALL-E 3, an AI text-to-image generator available to all Bing users for free, after rolling it out to select users about a week earlier. By October 5, reporting indicated the tool was already being used as part of a coordinated 4chan campaign “to flood the internet with racist images.”
In response, a Microsoft spokesperson told Ars Technica that the company was “investigating these reports and will take action as needed in accordance with our content policy, which prohibits the creation of harmful content.”
OpenAI also told the outlet that “the company prioritizes safety and has taken steps to limit DALL-E outputs.”
In addition to Bing’s content policies, OpenAI also prohibits use of its products for “hateful, harassing, or violent content.” Microsoft in November also announced a partnership with the group Tech Against Terrorism to combat “the use of digital platforms to spread violent extremist content” created by AI.
Numerous threads promoting Bing’s AI tool have appeared on 4chan’s far-right “politically incorrect” message board (known as “/pol/“) since the tool launched. Often titled “Memetic Warfare General,” the threads urge fellow users to create “propaganda” and note that “most people are using DALL-E 3.”
These threads often include a link to Bing’s image generator, a warning for other users to “be careful” as “certain buzzwords will eventually time you out,” and instructions to “SAVE ones you like, EDIT them to enhance them, and SHARE them later,” including “beyond” 4chan.
The threads have also included a link to a document with tips on how to “get around” being “censored by bing” and links to numerous “guides,” such as “The On Memewar Guide To Writing Better Propaganda,” a “good vs bad AI memes” guide, and a graphic on “the process” that guides users to save and edit the AI-created images and then share them on social media.
In response, users in those threads have shared numerous apparent AI-generated images they seemingly created, which often promote Nazi imagery and other far-right content such as Pepe the Frog. Some images have also featured Bing’s logo or pictures of singer Taylor Swift.
In total, Media Matters found more than 300 cases of users on “/pol/” featuring a link to Bing’s image generator since its initial launch in late September — despite the warnings to Microsoft and OpenAI about the message board’s activity in early October. The threads featuring the Bink image generator link in turn featured at least 100,000 replies, many of them including an apparent AI image.
Users in those threads have also boasted of creating “Pro-Jewish and Pro-Israel propaganda,” seemingly trying to exploit and inflame tensions over the ongoing conflict in Israel and Gaza (/pol/ is notoriously antisemitic). Some of the pro-Jewish and pro-Israel images that users have seemingly generated also include pictures of Swift.
One of the documents in the threads also mentions that they hope that the images can be collected in order “to produce these memes on our own machines without relying on Bing,” and some have claimed to have archived or promised to archive the images.
Earlier this year, users on /pol/ used an AI voice clone tool to create audio of celebrities using hate speech, forcing the tool’s creator company to announce “additional safeguards it will be introducing to the platform.” And a researcher trained another AI tool using millions of threads from /pol/, which was then used to post thousands of times on the message board.