Google's Veo 3 raises the floor on convincing fakes
When AI video crossed from obviously fake to genuinely hard to call.
- What it isA text-to-video generation model
- MakerGoogle DeepMind
- The riskConvincing misinformation at scale
- Surfaced2025 reporting
What changed
For years, the easiest way to spot AI video was simply to look: warped hands, melting backgrounds, audio that drifted out of sync. Google's Veo 3 narrowed that gap. It produces short, photorealistic clips with natural motion and synchronized sound from nothing but a text prompt, putting film-quality synthetic video within reach of anyone with a keyboard.
That capability is a creative gift and a misinformation problem at the same time. Reporting on the model's release warned how quickly believable scenes of events that never happened could now be produced and shared, and how thin the line between an impressive demo and a piece of disinformation has become.
Why it matters
Every leap in generation quality erases more of the visible artifacts that both detectors and ordinary viewers rely on. The historical fakes in this archive were often caught because something looked wrong. As that tell disappears, 'does it look real?' stops being a useful question.
What still holds up
When the pixels can no longer be trusted, authenticity shifts to two things that do not depend on visual perfection: provenance (cryptographic content credentials attached at creation, like C2PA and SynthID) and public signals (what the creator disclosed, what the platform labeled, and what informed viewers report). Provenance is powerful but fragile, because re-encoding by platforms often strips it. Public signals are what RealOrAiVideo reads, and they survive the upload.
Sources: TIME · The Verge (SynthID & C2PA). Further reading in the archive trackers.