May 20, 2026 Google says its SynthID system has already labeled more than 100 billion images and videos, along with 60,000 years’ worth of audio, as it ramps up efforts to track AI-generated content across the internet. In the early days of generative AI, spotting fake content was almost easy, as the extra fingers, warped faces, and glitchy artifacts gave it away without much effort. That reality has now flipped.
AI-generated images and videos have become so realistic that visual inspection alone is no longer reliable, forcing companies to build systems that can verify content at a deeper, technical level.
Google’s approach combines two methods. The first is SynthID, a watermark embedded directly into the content itself – inside pixels, video frames and audio waveforms. Unlike traditional labels, it survives edits like compression, cropping or rotation, making it harder to remove or tamper with.
According to Google DeepMind scientist Pushmeet Kohli, the system was deliberately designed to withstand attacks. The assumption was that any widely used detection method would be targeted, so robustness became a core requirement rather than an afterthought.
The second method relies on the C2PA standard, which attaches metadata describing how a piece of content was created. This is more transparent but also more fragile, since metadata can be stripped or altered as files move across platforms.
Google has already begun integrating C2PA into its products. Photos taken on Pixel 10 include details about edits and whether generative AI was used. The same capability is expanding to video on Pixel 8, 9 and 10 devices while Gemini is being updated to analyze and explain a file’s origin using this metadata.
Users can also upload content into Gemini and ask whether it is AI-generated. This works particularly well for media created using Google’s own systems, which already include SynthID watermarks.
Until now, that has been the system’s main limitation — coverage. SynthID only applied to Google-generated content, leaving a large portion of AI media untracked. That is starting to change.
Google has partnered with companies including NVIDIA, OpenAI, Kakao and ElevenLabs to bring SynthID into their own systems. The goal is to expand watermarking across a broader share of AI-generated content.
Even with that expansion, gaps remain. Many open-source and independently trained models still produce content without any watermarking, meaning detection will never be complete.
To make verification easier, Google is embedding SynthID checks across more tools. Features like Circle to Search, Lens and AI Mode will allow users to analyze content directly while Chrome will support Gemini-based scanning through shared tabs. A simple prompt like “Is this AI?” will trigger a detection scan.
At the same time, Google is limiting how broadly detection tools are exposed. There is no public API for SynthID yet, partly to avoid giving bad actors a roadmap for bypassing it. Instead, the company plans to release a controlled detection API through its enterprise platform for trusted partners.
The result is a layered system rather than a single solution. SynthID provides durable, embedded verification where it exists, while C2PA adds transparency where metadata can be preserved. Together, they represent an attempt to rebuild trust in digital content as AI makes the line between real and generated increasingly difficult to see.
