SynthID lands on Time's 2026 inventions list; 180 billion media files already carry its watermark

Google DeepMind's SynthID has been named one of Time's best inventions of 2026. More than 180 billion images and videos now bear its invisible watermark. The tool lets anyone verify whether a file was generated by AI. In a few years it has gone from an idea dismissed as futile to a de facto standard among the field's biggest players.

Roughly a decade ago — well before Veo 3 produced the viral clip of Will Smith eating spaghetti — Google considered building a detector a waste of effort. Pushmeet Kohli, DeepMind's vice president of science and strategic initiatives, explained that neural networks would eventually reach a point where they could no longer distinguish real images from synthetic ones. Rather than chase a detection capability that would only erode, the team chose the opposite approach: mark the content at the moment of creation.

SynthID embeds "small perturbations" in pixels — minute changes the human eye cannot perceive but the algorithm identifies with certainty. The mark survives cropping, compression and post-production editing. Download a video, trim it, re-upload at lower resolution, and the watermark remains readable. The technology is baked automatically into every output from Google's models, and also into models from OpenAI, Nvidia, Kakao, and soon Apple.

In October Google launched a public verification site. Any user can upload a file and get an instant answer on whether it carries SynthID. The move puts authentication in the hands not just of researchers and platforms, but of journalists, fact-checkers and ordinary users who encounter suspicious media online.

The broad rollout creates a new baseline: when a major model generates content, it is already "signed" by default. This does not solve fakes from open models or tools that have not adopted the standard. But it pushes the industry toward a state where the absence of a watermark becomes a warning signal in itself.