AI tools create misleading election images: Researchers

AI-based image generation tools such as OpenAI's DALL-E, Microsoft's Gemini, Stability AI's DreamStudio, and Midjourney, can produce misleading election-related images, according to researchers from the Center for Countering Digital Hate (CCDH), a nonprofit organisation.

According to these researchers, AI-generated images from these tools can contain voting-related disinformation, despite the companies behind the tools having specific policies that prohibit misleading content.

The Center for Countering Digital Hate (CCDH), a nonprofit that monitors online hate speech, used generative AI tools to create images of U.S. President Joe Biden laying in a hospital bed and election workers smashing voting machines, raising worries about falsehoods ahead of the U.S. presidential election in November.

"The potential for such AI-generated images to serve as 'photo evidence' could exacerbate the spread of false claims, posing a significant challenge to preserving the integrity of elections," CCDH researchers said in the report.

CCDH tested OpenAI's ChatGPT Plus, Microsoft's Image Creator, Midjourney, and Stability AI's DreamStudio, which can each generate images from text prompts.

The report follows an announcement last month that OpenAI, Microsoft, and Stability AI were among a group of 20 tech companies that signed an agreement to work together to prevent deceptive AI content from interfering with elections taking place globally this year. Midjourney was not among the initial group of signatories.

CCDH said the AI tools generated images in 41% of the researchers' tests and were most susceptible to prompts that asked for photos depicting election fraud, such as voting ballots in the trash, rather than images of Biden or former U.S. President Donald Trump.

ChatGPT Plus and Image Creator were successful at blocking all prompts when asked for images of candidates, said the report. However, Midjourney performed the worst out of all the tools, generating misleading images in 65% of the researchers' tests, it said.