In today’s internet, AI-generated images, videos, and voices are nearly indistinguishable from human-made content. Deepfakes, synthetic audio, and manipulated visuals spread faster than fact-checkers can keep up. The question is no longer if you’ll encounter AI-made media, but how you’ll know when you do.

Enter SynthID
Google DeepMind’s SynthID is a digital watermarking technology designed to embed invisible signals into AI-generated media. These watermarks are imperceptible to the human eye or ear but remain detectable even after edits like resizing, compression, or screenshots. They act as a hidden signature, traveling with the file wherever it goes.

So far, SynthID has already been applied to over 100 billion images and videos and the equivalent of 60,000 years of audio. That scale makes it one of the most widely deployed provenance tools in the AI space.

New cross-platform coverage
With OpenAI, Kakao, and ElevenLabs joining the initiative, content from ChatGPT, Kakao’s services, and ElevenLabs’ voice tools will carry the same SynthID watermark used across Google’s products. The goal is to provide a consistent signal that content is machine-generated, even after common edits.

Integration with search and browsers
Google also said it is expanding provenance tools based on C2PA Content Credentials into the Gemini app, Google Search, and Chrome. These credentials help users learn whether a piece of media was captured by a camera, generated by AI, or edited with AI tools. SynthID verification is already available in the Gemini app and has reportedly been used 50 million times worldwide. Google plans to add verification to Search immediately and to Chrome via Lens, AI Mode, Circle to Search, and right-click tools in the coming weeks.

The moves aim to make it easier for people to assess the origins of online media at a time when deepfakes and AI-generated content are becoming harder to distinguish from human-made material. By aligning major model providers and platform vendors behind a shared watermarking approach, the industry hopes to establish a practical, interoperable way to signal provenance.

Experts caution that watermarks are not a silver bullet. Techniques such as heavy editing, format conversion, or re-generation can remove or obscure signals. Researchers have also shown that determined actors can sometimes evade detection through methods like paraphrasing or re-rendering content. Adoption by more providers reduces friction for verification but does not eliminate misuse entirely.

Broader adoption of SynthID and Content Credentials could standardize how services label AI-created media, helping platforms, publishers, and users make quicker, more informed judgments about authenticity. The coming weeks will test how reliably the system works in the wild and whether adversaries find practical ways to bypass it.