Online scam videos pose serious threats to individuals and society. Existing work mainly treats the problem as binary classification, offering little insight into scam mechanisms, struggling to separate videos that look like ads or other benign content, and lacking robustness against misleading arguments. To address these gaps we introduce the MechScam benchmark, which contains 6,614 videos spanning diverse scam mechanisms and non‑scam categories. Each entry includes evidence‑backed explanations, annotations of potential scam elements with context, and deliberately crafted misleading arguments that misuse genuine evidence. We also propose the ScamCourt debate framework: trained prosecution and defense debaters present opposing arguments while a trained judge decides based on a transcript enriched with verifiable citations. Extensive experiments show that state‑of‑the‑art multimodal LLMs fail to reliably distinguish scam videos from look‑alike non‑scam content. ScamCourt outperforms larger LLMs and same‑backbone fine‑tuned baselines, reducing false positives and showing stronger resistance to deceptive reasoning. Together, MechScam and ScamCourt advance reliable and interpretable scam video detection. The dataset and code are publicly released.\n\nReview