From Manipulation to Mistrust: Explaining Diverse Micro-Video Misinformation for Robust Debunking in the Wild
A new AI framework tackles 10,000 real-world fake videos, outperforming existing models by analyzing multimodal manipulation.
A research team led by Zhi Zeng has published a significant paper, "From Manipulation to Mistrust," accepted at WWW 2026, that tackles the complex problem of micro-video misinformation. They identify a critical gap: existing benchmarks and models typically focus on a single type of deception, failing to capture the real-world diversity of fake content, which includes multimodal manipulation, AI-generated clips, exploitation of cognitive biases, and out-of-context reuse of real footage.
To address this, the team built WildFakeBench, a large-scale, realistic benchmark comprising over 10,000 real-world micro-videos. Each video is annotated with expert-defined labels that attribute the specific type and method of misinformation, moving beyond simple binary "real/fake" classification. On this foundation, they developed FakeAgent, a novel multi-agent AI framework inspired by Delphi's reasoning processes. FakeAgent doesn't just detect; it explains by jointly analyzing video content and retrieved external evidence to pinpoint the manipulation technique.
Extensive experiments demonstrate that FakeAgent consistently outperforms existing multimodal large language models (MLLMs) across all misinformation types covered in WildFakeBench. The benchmark itself is designed as a challenging testbed to advance the field toward more explainable and attribution-grounded detection systems. The researchers have made both the WildFakeBench dataset and the FakeAgent code publicly available, providing essential tools for the community to build more robust debunking systems.
- WildFakeBench is a new benchmark with over 10,000 real-world micro-videos, each annotated for specific misinformation types and sources.
- FakeAgent, a multi-agent reasoning framework, outperforms existing MLLMs by integrating multimodal analysis with external evidence retrieval.
- The system provides fine-grained attribution, explaining *how* a video is misleading—whether via AI generation, context manipulation, or cognitive bias—not just labeling it as fake.
Why It Matters
This provides a scalable, explainable framework for platforms and fact-checkers to combat the next generation of AI-powered video disinformation.