TextFake: Benchmarking AI-Generated Image Detection on Text-Rich Images
Researchers have developed a new benchmark called TextFake to evaluate the effectiveness of AI-generated image detection systems on images containing text. Existing detectors perform poorly on these text-rich forgeries, such as fake screenshots and documents, with accuracy dropping significantly compared to natural images. The benchmark, comprising 20,000 images across 28 languages, reveals common failure modes including issues with text density, rendering fidelity, and sensitivity to minor perturbations. AI
IMPACT Highlights critical vulnerabilities in AI image detection, potentially impacting misinformation efforts and requiring new detection methods for text-heavy fakes.