Scene A04:12:08
Generated
Move / touch to inspect
ICLR 2026 submissionVideo forensics · Social risk
Can we defend
against generated
crisis videos?
A systematic evaluation of detectors, generators and social dissemination—anchored in footage from real-world crisis events.
Synthetic footage / generated frame
Move across the image to disturb the evidence.
17,886Total videos
1,830Real anchors
16,056Generated clips
9Video generators
10Social-risk categories
01 The question
Realism is moving faster than our ability to verify it.
Modern video generators can fabricate coherent depictions of wars, disasters, public emergencies and other high-consequence events. Yet existing benchmarks provide limited evidence about how detectors behave when the footage looks plausible, circulates socially, and is anchored to a real event.
RA-Bench closes that gap by pairing real crisis footage with generated continuations made from the same first frame and prompt—then testing machines, people and the transformations that happen during sharing.
02 Perception test
HALFTONE REVEAL / FUEL TANK FIRE
Can you tell real from generated?
Scene B04:12:08
Real
Move / touch to inspect
Both clips depict the same event. Which one is generated?
03 Collaboration
Built across disciplines,
institutions & borders.
Thirty-five contributors across nineteen affiliations bring together video generation, multimedia forensics, multimodal reasoning and responsible AI.
* Equal contribution · § Project lead · † Corresponding authorsView affiliation artwork
04 The benchmark
Same event.
Different reality.
RA-Bench uses real videos as anchors. Each generated continuation inherits the scene context—forcing detectors to reason beyond obvious content mismatch.
Generated video atlas03 scenarios · 15 moments · hover to isolate
Collect
675 public-source videos grounded in real, socially consequential events.
Segment
Scene boundaries, duplicate filtering and encoding unification.
Review
Two-stage human screening for relevance, quality and rights metadata.
Anchor
1,830 standardized real clips across 10 risk categories and 44 subcategories.
Generate
16,056 paired clips from four open and five closed generators.
05 What we found
Three failure modes.
One urgent signal.
01Generalization gap
43.9–57.3%
source-level mean AUC
Traditional detectors drop sharply from public-reference AUCs of 67.6–98.6%, and their rankings shift across generators.
02HumanProof
633
generated videos called real ×5
Every clip in this subset fooled all five reviewers. Traditional detectors average just 47.5% AUC on the resulting challenge set.
03LastMile
46.0 → 1.4%
fine-tuned MLLM mean FakeR
A realistic chain of platform transformations shifts detector predictions toward real, even when the underlying event content is unchanged.
06 RA-Bench-LastMile
Every share leaves a trace.
Detectors lose the trail.
RA-Bench-LastMile simulates a sequential dissemination chain. Platform transformations preserve the scene for people while progressively weakening the signal available to detectors.
Original46.0%
DownloadT1
MessengerT2
Screen recordT3
Re-uploadT4
Full chain1.4%
07 Read the research
Detection must survive
the real world.
Explore the complete benchmark construction, detector protocols, generation analyses, human study and dissemination experiments.