AI-generated images tied to news events and online claims can mislead audiences, enable fraud, or spark public controversy. Detectors designed to spot such images are often evaluated on curated datasets, which may not reflect how these images appear in the wild.
To address this, the paper introduces WildAIGC-Bench, a benchmark built through recurring collection of images from news websites and fact-checking sources. It includes both fully generated and edited images that have actually circulated online, providing a more realistic test bed.
By testing detectors on this material, the benchmark aims to show how well they hold up in practical settings. The work is described in a preprint on arXiv, and no independent evaluation of the benchmark's claims is included in the source.