SynthID

Today, computer programs can generate images, music, and text that are barely distinguishable from human work. This raises a practical question: how can one later tell what came from such a program? SynthID is Google's answer to this. It is a digital watermark, that is, a hidden marker that is built directly into the file during generation. It is invisible or inaudible to humans, but a special verification program can find it again. Google has been using SynthID in its own services since 2023 and, according to its own figures, has by now marked billions of pieces of content with it.

Why proof of origin suddenly became a topic

In the past, a photograph was a reasonably reliable piece of evidence that something had really happened. That certainty has disappeared. A convincingly faked image today costs neither money nor skill. Faked recordings of politicians, invented disaster photos, and cloned voices have already appeared in election campaigns. When no one can tell the difference anymore, eventually even genuine material is doubted.

On top of that comes pressure from lawmakers. The European Union’s AI Act requires that artificially generated content be labeled in a machine-readable way. Providers of large AI services must therefore present a technical solution. SynthID is one of the proposals for what that could look like.

For schools and universities, this sounds like a convenient answer to copied homework. That hope should be tempered. SynthID only detects content that comes from systems that actually build it in. Anyone using a different program leaves no trace. A negative verification result therefore proves nothing.

What exactly is hidden in the file

For images, SynthID minimally alters individual pixels. The deviations are so small that the eye does not notice them, but they follow a fixed pattern across the entire image. Importantly, the marker is not stored in the file’s metadata, which could simply be deleted. It is embedded in the image itself. That is why it usually survives cropping, compression, or color changes.

For text, it works differently, because letters cannot be shifted slightly. A language model selects each word from a list of possible continuations, which it rates with probabilities. SynthID shifts these ratings slightly in favor of certain words according to a secret pattern. Over a longer text, this creates a statistical signature. The verification program calculates how strongly a text follows this pattern.

This is precisely where the limits come from. Short texts contain too few word choices, so the signal becomes too weak. Anyone who heavily rewrites the text or runs it through a translation program often destroys the pattern. And the result is never a simple yes or no, but a probability.

SynthID in Google’s products and in the debate

All images from Google’s image generator Imagen and from the video model Veo are marked, as is music from the audio model Lyria. Responses from the chatbot Gemini also carry the text marker. In Google Image Search, there is a feature called 'About this image' that shows when an image contains a SynthID label. Since 2025, files can also be uploaded by users themselves to a verification portal.

In the news, SynthID mostly comes up in connection with deepfakes, that is, deceptively realistic fakes of people. Google has released parts of the technology as open source so that other companies can adopt it. In parallel, an industry alliance called C2PA is working on a different approach: there, provenance data is attached to the file as a kind of digital package insert. Both approaches are not mutually exclusive and are sometimes combined.

A common misconception is that SynthID is a detector for AI content in general. It is not. It is proof of content from certain systems, comparable to the watermark on a banknote. It shows that a particular note is genuine, but it reveals nothing about notes from a different printing press.