Yemen's Houthi rebels struck Saudi Arabia's King Khalid International Airport multiple times since last Thursday, killing at least 16 people. Real footage exists. But DW Fact Check found that four of the most viral videos circulating on social media — collectively viewed millions of times — were AI-generated fakes, not authentic documentation of the attacks. The fakes range from crude to sophisticated. One video, online since March 2026, shows people running toward an explosion rather than away from it, with hands appearing and disappearing between frames. Another was reverse-image-traced to a 2016 Shutterstock photo of the airport, used as a seed image for an AI video generator that added a missile strike — complete with a vehicle that changes shape mid-explosion and shadows that don't match the fireball. A third shows people walking through queue barriers as if the objects don't exist. The fourth carries Google's SynthID watermark in both image and audio tracks, confirming it was made with Google AI tools. The distribution footprint is the real story. These videos appeared in Arabic, English, Russian, Spanish, Turkish, Nepali, Greek, and Albanian. They circulated on X, YouTube, Instagram, TikTok, and Facebook. The Times of Israel shared one as authentic, racking up 256,000 views. The most-viewed single post hit 2.9 million views on X alone. Combined reach across platforms likely exceeds tens of millions of impressions before any correction. Detection tools exist but arrive too late. Meta's AI identification tool flagged one video. Google's SynthID watermark caught another. Frame-by-frame analysis by human fact-checkers caught the remaining two through visual anomalies — morphing limbs, impossible physics, blurred glitches. But every one of these interventions happened after mass distribution. The EU's AI Act has required machine-readable markings on AI content since August 2026, yet the watermarked video still went viral on TikTok without any platform-side interception. The structural pattern is clear: AI-generated conflict footage is now trivially cheap to produce (all four videos were under 15 seconds — the sweet spot for generative AI), nearly free to distribute at scale, and expensive and slow to debunk. The asymmetry runs entirely in favor of the fabricator. A single account identified by DW as a repeat AI-content producer posted one of these videos on Facebook, where it was correctly labeled as AI content — yet the same video circulated unlabeled on other platforms. The cost structure tells you where this goes. Generating a convincing 10-second clip costs pennies and minutes. Fact-checking one requires reverse image searches, frame-by-frame analysis, watermark detection tools, and multilingual monitoring — hours of skilled labor per video. At current ratios, the fabrication side scales exponentially while the verification side scales linearly at best. Every future military conflict will now include a synthetic media layer that most viewers cannot distinguish from reality. The regulatory and platform response remains fragmented. The EU mandates watermarks; Google complies with SynthID; Meta labels AI content on Facebook. But none of these mechanisms prevented mass distribution of fakes across the platforms where most people actually encountered them. The watermark worked exactly as designed — and was completely irrelevant to the outcome.
