AI image generators can make pictures quickly, but the tools can also be repurposed to produce racist content for circulation online. A 404 Media report described a 4chan thread that recommended image tools and gave users instructions for adding captions and sharing the results across social platforms.
A guide to making and spreading images
The thread named several options, including Stable Diffusion and DALL-E, and linked to Bing AI’s text-to-image generator as a “quick method.” The instructions also pointed to conventional image editing software such as Photoshop, showing that the suggested campaign did not depend on a single generator.
Users were told to add provocative captions and post the images on social media. The guide described captions as “funny, provocative” and recommended messages that were easy to understand. Its examples included a claim about Jews and 9/11.
One image in the guide showed a crying Pepe the frog beside a needle, with a gun pointed at his head. The suggested caption was “vaccines enforced by violence.” Another example depicted two Black men with gold chains chasing a white woman and recommended adding a “redpilling message.”
Bing appeared often in the examples
404 Media assessed that many images in the thread appeared to come from Bing. The report pointed to their square format and uniform 1024 x 1024 resolution as clues. It said images were then shared on Telegram, X (formerly Twitter), and Instagram.
Those details suggest how image generation can fit into a broader distribution effort: a person can create or edit an image, attach a message intended to provoke, and post it where audiences may encounter it. The report describes a specific thread and examples; it does not establish how widespread this activity is beyond them.
Filters can be tested through wording
In a test attempting to recreate an example from the guide, 404 Media found Bing rejected the prompt “two angry Black men chasing a white woman.” It accepted “photorealistic two angry Black rappers chasing woman.” The different outcomes illustrate a challenge for image filters: a changed prompt may still seek a harmful depiction while avoiding wording that a system rejects.
The report said it was unclear what steps the makers of the image tools had taken to block the methods used to get around filters. An OpenAI spokesperson told Ars the company prioritizes safety and has sought to limit harmful DALL-E outputs, including images requested using a public figure’s name. The spokesperson also said Microsoft applies its own safeguards for DALL-E 3.
Safety measures face a difficult test
Microsoft told Ars that Bing Image Creator was designed to support creativity, but that some users were trying to use it in unintended ways. The company said it was investigating the reports and would act as needed under its content policy, which prohibits harmful content. It also described work on techniques and safety systems guided by its responsible AI principles.
Stability AI did not immediately respond to Ars’ request for comment. The reporting also recalled earlier concerns about racist and sexist biases in image generators. OpenAI had said in comments to Vice in 2022 that it used curated training data, filters, and human and automated monitoring to mitigate harmful outputs, and that it was working to measure and address biases.
The issue raised by the 4chan thread is twofold: image systems can reflect bias, and people can deliberately steer them toward offensive results. Safeguards may block some requests, as Bing did in 404 Media’s test, while a reworded prompt can produce a different outcome. The report leaves open how Microsoft, OpenAI, and other AI companies will respond as users find ways to put these tools to harmful use.