Found first: a primary source the press has not covered yet.
A 35-author team led by Shuo Liang has released RA-Bench, a benchmark showing that no current AI video detector reliably identifies synthetic depictions of wars, disasters, and public emergencies. Their paper, "Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events?", submitted August 14, tests 19 detection systems across 17,886 videos and finds that social media dissemination compounds the failure.
What the source says
The researchers built RA-Bench from 1,830 real-world anchor videos spanning 10 social-risk categories, then generated 16,056 synthetic clips using nine video generators, four open-source and five closed-source. Against this dataset they ran seven traditional detectors, ten zero-shot multimodal models across three review settings, and two multimodal large language models fine-tuned specifically for the task. None of the three detector families generalizes consistently across benchmark instances. Videos that mislead human viewers are also difficult for the detectors, and social dissemination substantially reduces detection reliability further.
Why it matters
Crisis-event video is where misinformation does the most damage: it shapes emergency response, public trust, and political decisions in real time. The failure is not confined to one model family. It holds across 19 systems representing three distinct approaches to the problem. The social-dissemination finding adds a compounding factor: the videos most likely to spread are the ones hardest to catch. Researchers now have a structured test bed for this specific failure mode; anyone relying on the assumption that detectors can screen crisis footage has a concrete counter-example.