A crisis clip can reach a fact-checker after compression, resizing, a frame-rate change, and a badge have altered the file. A detector that flagged the original may call the delivered copy real.
What changed
A new preprint introduces RA-Bench, a benchmark of 17,886 videos across ten broad social-risk categories. It pairs 1,830 real-event anchors with 16,056 generated clips from nine video generators. The authors evaluated seven traditional detectors, ten zero-shot multimodal models, and two fine-tuned multimodal models under several review settings.
This is where the delivery problem gets concrete. For each condition in a controlled last-mile simulation, the authors used 150 real videos and 1,350 matched generated clips. They applied common changes to both groups: transcoding, half-size downsampling, a reduction to eight frames per second, a synthetic news badge, and a sequence combining all four.
The result was uneven but severe. Under the full sequence, mean fake recall across five fine-tuned configurations fell from 46.0 percent to 1.4 percent. Across seven traditional detectors, mean AUC fell 4.2 percentage points, although individual detectors behaved differently. The badge alone pushed the fine-tuned configurations toward a Real judgment even though the underlying scene did not change.
Why it matters
That shift exposes a basic weakness in single-score verification.
A detector score belongs to a particular version of a file. Platform delivery can change codec, size, frame rate, and overlays before a reviewer sees the material. Testing only the original or only the received copy leaves part of the chain unexamined. So the safer operating model is layered: preserve the best available original and its source trail, record each transformation, and test the version that viewers actually receive. Keep detector output beside provenance, chain-of-custody evidence, and accountable human review rather than allowing one score to settle a high-consequence question.
Watch next
But the limits are substantial. The paper is a v2 preprint and the last-mile test is a simulation, not a study of actual platform circulation. It removes audio and metadata, comes from the benchmark authors, and has not been independently reproduced here. Its results do not establish how common detector failures are in the wild.
Independent replication would strengthen the finding, as would tests against unseen generators and measurements taken from real platform delivery paths. Until then, one practical check is available: ask whether the detector saw the evidence people received or a cleaner file that never reached them.
Primary source: Liang et al., RA-Bench, arXiv:2608.14391v2