EVID-Bench: When Seeing Is No Longer Believing—A New Benchmark for Search-Driven Video Misinformation Detection

This article introduces EVID-Bench, a benchmark for search-driven video misinformation detection. Targeting covert video manipulations at the semantic and evidential levels (such as selective clipping, AI-generated content injection, etc.), the benchmark requires models to actively search for relevant videos on the open web and identify misinformation through cross-video comparison. The benchmark includes 222 video samples covering 9 manipulation types. Existing state-of-the-art multimodal models perform poorly on this benchmark, highlighting the need to build intelligent systems with active search and cross-source verification capabilities.

In the era of information explosion, video misinformation spreads rampant. Traditional detection focuses on pixel-level tampering (e.g., Deepfake), but more covert and dangerous manipulations occur at the semantic and evidential levels: real materials are selectively clipped, time-rearranged, cross-source spliced, or injected with AI-generated content to construct false narratives. Such manipulations cannot be judged as true or false by humans or advanced AI models based solely on the video itself, as the missing evidence is not inside the video.

EVID-Bench (Evidence-based Benchmark) is a search-driven video misinformation detection benchmark. Its core elements include:

• 222 video samples: covering various sources and topics
• 9 manipulation types: divided into three categories (AI-generated, single-source editing, multi-source splicing) Key feature: All samples cannot be detected by state-of-the-art models through visual inspection alone; models need to understand context, retrieve external evidence, and perform logical reasoning.

The research team evaluated 9 state-of-the-art multimodal models, using retrieval-augmented verification as the baseline:

• Best system accuracy: 61.43% at the point level, 43.24% at the video level
• AI-generated manipulations are particularly difficult to detect, as their visual quality is hard to distinguish from real videos
• Typical error patterns: fixation on irrelevant anchors, misattribution of synthetic content, premature termination of search

Implications for the AI Research Community

• Beyond end-to-end thinking: need to integrate external knowledge retrieval
• New challenge for multimodal reasoning: shift from passive viewing to active investigation
• Expansion of the RAG paradigm: extend to verification and fact-checking fields

Implications for Misinformation Governance

• New battlefield for technical confrontation: rely on cross-source verification
• Platform responsibility: need to establish cross-video retrieval and comparison mechanisms

Limitations and future directions of EVID-Bench:

• Real-time challenge: Practical applications require extremely short time to complete search and comparison
• Multilingual and cross-cultural: The current benchmark is mainly based on English content
• Adversarial evolution: Misinformation creators will adjust their strategies to counter detection technologies

EVID-Bench reminds us that 'seeing is believing' no longer holds in the era of rampant AI-generated content. Building intelligent systems with active search, cross-source comparison, and logical reasoning capabilities is key to addressing the next generation of video misinformation. This benchmark provides researchers with an evaluation tool and points the way for the industry—from pure content understanding to evidence-driven intelligent verification.