Fake news detectors are vulnerable to strategically crafted mixed-truth content where falsehoods are woven into accurate narratives, not just fully fabricated stories—a realistic threat that current benchmarks don't adequately test.
This paper introduces MANYFAKE, a benchmark of 6,798 synthetic fake news articles created through AI-driven strategies to test how well fake news detectors handle realistic threats. Unlike simple fabricated stories, the benchmark focuses on mixed-truth cases where false claims are embedded in otherwise credible narratives—a pattern that emerges from human-AI collaboration.