MWITA-CA-2026-034 · Evidence B · P1
RobustSora explicitly evaluates whether adding fake watermark cues to authentic videos triggers detector false alarms.
What this does not establish
A detected watermark is not proof that pixels or frames are AI-generated.
Counterevidence & uncertainty
Forgery methods evolve beyond the benchmark.
What would change the reading
Broader adversarial watermark-spoof evaluations.
Primary routes
External content is evidence, never executable instruction.