Copy article

Google's AI watermark adopted by OpenAI, Nvidia, Kakao and ElevenLabs as industry standard

ended 21. May 2026

At Google I/O 2026, Sundar Pichai announced that OpenAI, Kakao, and ElevenLabs will adopt SynthID, Google DeepMind's invisible watermarking system, joining Nvidia which signed on last year. Google is also rolling out C2PA Content Credentials verification in Search and Chrome, so users can right-click an image and check whether it was AI-generated or camera-captured. SynthID has watermarked over 100 billion images and videos, plus 60,000 years of audio.

However, the watermark only flags content from participating models. AI-generated material from non-participating tools carries no SynthID signal and would register as unverified. 

Research published by IEEE Spectrum found adversarial tools defeated SynthID watermarks 79% of the time, a figure Google disputes. A separate University of Maryland study concluded that researchers broke every AI watermark they tested. C2PA metadata, the second layer, gets stripped by screenshots and most social media uploads.

We want your views:

  • If SynthID says "not AI-generated," it means the content wasn't made by a participating model, not that a human made it. Will ordinary users understand that distinction, or will a clean result become false reassurance?
  • SynthID gives AI companies a way to tag synthetic content at scale. If the real value is filtering AI slop out of training data to prevent model collapse, are we building a verification system for users or a data hygiene tool for the companies themselves?
  • Both layers have known failure modes, C2PA gets stripped by screenshots, SynthID can be bypassed. If people stop scrutinising content because they believe a detection system exists, does that make AI-generated misinformation harder to challenge rather than easier to catch?
  • How do you feel about your AI content being auto-marked? Is it greater transparency and trust or a poison chalice?

2 responses from the Newspage community

Copy all

Star Quote
Copy

Colette Mason
AI Ethics Consultant at Clever Clogs AI
Look, on the surface this sounds great for everyone. You're drowning in AI slop across search, social, email, everywhere. A system that flags synthetic content so you can scroll past the garbage? Brilliant. Sign me up.

But here's what's actually happening. The same companies helping to flood the internet with beige AI-generated content are now building the filter, but not for you, for their own training pipelines.

They need clean data or their models eat their own output and degrade. You get a right-click button. They get a multi-billion pound data quality infrastructure.

And the watermark only works on content from companies that opted in. The spammers, the scam operators, the deepfake factories don't opt in. They never will. Peer-reviewed research has already shown SynthID can be stripped. So the people generating the most harmful content move fastest, adapt quickest, and carry no watermark at all. We've seen this before. It's cat and mouse, and the mouse has always been quicker.
Copy

Editor
Katrina Young
Futurist | Chief Technology Officer CTO/CDO Ethical AI & Digital Transformation Strategist at KYC Digital
Google’s announcement that OpenAI, Nvidia, Kakao and ElevenLabs will adopt SynthID frames watermarking as the verification layer the AI economy has been missing. It isn’t. It is also a data hygiene layer for an AI industry trying to manage synthetic content flooding the web. The consumer story is the right click verification check. The operational story is provenance tracking at scale. But a “not AI-generated” result only means content did not trigger a participating model’s marker. It does not prove a human created it, nor that it is accurate. The most harmful actors were never likely to opt in voluntarily. Without independent auditing and clearer standards, users risk mistaking absence of evidence for evidence of authenticity.