4chan users are orchestrating a posting campaign in which they utilize Bing’s AI text-to-image generator to craft racist images, which they subsequently disseminate across the internet. This development highlights how users can manipulate freely accessible, user-friendly AI tools to swiftly inundate the internet with offensive content, even when those tools are supposedly under strict moderation.
The 4chan thread encourages participation in this effort, framing it as a form of propaganda for amusement, and urges others to join by creating, editing, and sharing such images. A visual guide hosted on Imgur, linked within the post, provides instructions for users to employ AI image generators, customize them with captions resembling political campaigns, and then post them on social media platforms, particularly Telegram, Twitter, and Instagram. It has been observed these images shared on a TikTok account, which has since been removed.
The guide offers flexibility in methods, including the use of Stable Diffusion, an open-source AI model allowing unrestricted image generation, or the well-known Photoshop. However, it notes that most users are utilizing DALL-E 3 through a provided link to Bing, labeled as the “QUICK METHOD.”
Based on the images’ default square format, uniform 1024 x 1024 resolution, and the recent notoriety of the 9/11 trick, it appears that most images in the thread are generated using Bing.
The visual guide instructs potential posters to create images that are humorous, provocative, and convey certain messages, such as implicating Jews in 9/11 or highlighting migrant crime in Scandinavia. Notably, my tests on Bing showed that while it would reject prompts like “two angry Black men chasing a white woman,” it would generate an image closely resembling the one shared in 4chan’s guide when the prompt was more nuanced.
As the 4chan thread acknowledges, individuals have used Photoshop and less advanced tools to create and disseminate racist images on the internet for years. The issue of racism is not solely a technological one, but it’s crucial to recognize that technology is now amplifying racist content creators, enabling them to produce more sophisticated material more rapidly than previously observed online. Moreover, they are doing so with tools that are ostensibly “safe” and rigorously moderated, even to the extent that they won’t generate completely innocuous images like those of Julius Caesar. Consequently, we find ourselves facing the worst of both worlds from Bing—an AI tool that refuses to generate certain content but is inadvertently bolstering 4chan’s racist efforts.
This underscores a point made by Sam yesterday, that most generative AI models with strict filters and terms of use, whether generating images or text, engage in a never-ending game of circumvention. Microsoft can ban specific phrases from prompts indefinitely, but individuals will always find ways to bypass filters.
The distinction here is that these images are explicitly racist. In less than a day, we’ve witnessed an AI image generator loophole being exploited to conduct a coordinated, explicitly racist posting campaign, reshaping the broader internet to resemble the most unpleasant aspects of 4chan.