Can Google’s SynthID watermark really curb AI misinformation?

Google has released a technical report on SynthID, its invisible watermarking technology meant to identify content generated by its models (Gemini, Imagen, Lyria). The verdict from independent tests: the tool holds up well against common manipulations — cropping, compression, color changes. But researchers at the Starling Lab, a Stanford-USC research collective specialized in digital content authenticity, are notably tempering the enthusiasm.

SynthID works, but only covers part of the problem

The principle is simple on paper: a watermark imperceptible to the naked eye but detectable by software, embedded directly into an image's pixels or into the word probabilities chosen by a text model. Technically, SynthID has proven itself: it remains detectable even after fairly aggressive modifications.

The problem, according to Adam Rose, an advisor at the Starling Lab, doesn't come from the technology itself. It comes from everything that lies outside its scope: open-weight (open-source) models, which can run directly on someone's personal computer without going through any watermarking infrastructure. No law, no industry standard can force a locally run model to self-mark.

In other words: SynthID could very well become the standard among major publishers — Google, and potentially their partners — without ever covering the real majority of content produced worldwide.

AI and copyright: what creators need to understand

Publication date: May 27, 2026 I Reading time: 9 minutes 30s

The real risk: a false sense of security

This is the point that deserves the most attention, and it echoes a concern we share internally. A watermark answers the question "was this content generated by a tool that watermarks its output?" — not the question "is this content true?" These are two different questions, and the risk lies in conflating them.

In practice: a reader, customer, or partner used to the idea that AI content is "marked" could develop excessive trust in any unmarked content — even though the absence of a watermark proves nothing at all, neither human origin nor accuracy.

There's also a less commonly discussed perverse effect. As these standards become widespread, the use of AI in content production becomes an argument competitors can turn against a brand, whether or not it actually uses AI. And audiences are increasingly able to spot, by eye, a "too polished" text or a "too perfect" image — with or without a technical watermark. Betting on a watermark to manage your reputation, rather than on the quality and transparency of the published work, is a risky bet.

What this changes for anyone working with AI day to day

The good news in all this: the solution isn't to hide AI use, nor to rely on a watermark to do it for you. It's to keep control over what AI produces — verify, cross-check, correct before publishing — and to choose the right model for the right task rather than depending on a single closed ecosystem.

What to Remember

• The SynthID watermark is a genuine technical advance, but it won't solve disinformation on its own: too much AI content will structurally escape its coverage, and marked or unmarked content says nothing about its accuracy.
• Human vigilance — verifying, cross-checking, comparing several sources and several models — remains, and will remain, the best protection.
• That's also why we believe it's better to own and control your AI use than to rely on a watermark to decide for you.