AI image detectors can be easily fooled, new report finds

• Adding grain to AI-generated images makes it harder to identify fakes, The New York Times reports.
• The probability of detection drops from 99% to 3.3% when pixelated noise is added to the image.
• The findings come as users in the U.S. and internationally begin to use AI imagery to influence election campaigns.

From faked campaign ads to stolen works of art, AI-generated images have been responsible for a wave of online disinformation in recent months.

Now, The New York Times reports that AI detection software (one of the first line of defense against the spread of AI-generated disinformation) can be easily fooled simply by adding granularity to AI-generated images. reported.

A Times analysis found that when editors add grain, or texture, to AI-generated photos, the likelihood of software identifying the image as AI-generated drops from 99% to just 3.3%. To do. Even the software Hive, which had the highest success rate in the Times report, failed to correctly identify AI-generated photos after editors pixelated them further.

As a result, experts warned that detection software should not be the only line of defense for companies trying to combat misinformation and stop the distribution of these images.

“Every time someone builds a better generator, people build better discriminators, and then people use better discriminators to build better generators,” said Duke University Computers. Science and engineering professor Cynthia Rudin told The Times.

The paper’s analysis comes at a time when users are increasingly deploying AI-generated misinformation online to influence political campaigns, insiders said. For example, Ron DeSantis’ presidential campaign earlier this month distributed fake images of Donald Trump and Anthony Fauci.