What is SynthID?

SynthID is Google DeepMind's watermark for AI content from Google's models, covering images, audio, video and text. Unlike Content Credentials, which attach a signed record to a file, SynthID alters the content itself, so it survives screenshots, cropping and compression. A detection tells you content came from a model that applies it; no detection tells you nothing at all.

Where it sits

SynthID is Google DeepMind's watermark for content its own models produce. Images from Imagen, video from Veo, audio, and text from Gemini all carry it. Google has said it has marked well over a hundred billion pieces of content this way, which makes it the largest deployment of anything like it.

The mark is imperceptible by design. Nothing about a SynthID image looks different to a person, and no visible badge is added. It is only recoverable by a detector that knows what pattern to look for.

Why it is not metadata

This is the distinction that matters and the one most explanations skip. Content Credentials, the C2PA standard, work by attaching a signed record to the file. SynthID changes the content itself, nudging pixel values (or audio samples, or word choices) into a pattern carrying the signal.

The consequence is durability. Everything that destroys attached metadata leaves SynthID alone. Screenshot a SynthID image and the watermark is in the screenshot. Crop it, compress it, re-save it in another format, and it is still detectable. That is the specification, not a happy accident, because a watermark that broke under ordinary handling would mark nothing worth marking.

Two schemes, opposite failure modes

Content Credentials (C2PA)

A signed record attached to the file
Carries who made it and every edit since
Tampering breaks the signature, so it shows
Destroyed by a screenshot, a re-encode, or most uploads

SynthID

A pattern inside the pixels themselves
Carries only that a marking model made it
Nothing to verify against, you need the detector
Survives screenshots, cropping and compression by design

Watermarking text is the harder case

Images have somewhere to hide a signal. Text does not, and SynthID's approach for text works differently: it biases the model's word choices as it generates, in a statistical pattern a detector can spot across a long enough passage.

It is inherently weaker than the image version. Short passages carry too little signal, and paraphrasing a piece of text substantially degrades it. Treat text watermarking as much less reliable than the image case, because it is.

What it does not tell you

A SynthID detection says the content came from a model that applies SynthID. It does not tell you that content without one is real. Most generated content on the internet comes from tools that mark nothing at all, so an unmarked file is uninformative rather than reassuring.

This asymmetry is the single most misunderstood thing about the whole category. Presence is evidence. Absence is silence. Any tool giving you a confident percentage that an unmarked file is human-made is guessing and dressing the guess in a number.

Can it be removed

Not by anything in a browser. The tools marketed as AI watermark removers strip metadata, which removes Content Credentials and does nothing to SynthID, since the two live in completely different places. A screenshot achieves the same metadata removal for free.

Determined adversarial attacks on pixel watermarks are a live research area and no scheme claims to be unbreakable. That is a different statement from re-saving a file, which is what the consumer tools do.

The related question, what actually survives, covers the three layers and how each one fails.

Common questions

What is SynthID?
Google DeepMind's watermark for AI content produced by Google's own models, covering images, audio, video and text. It alters the content itself into a pattern a detector can recover, rather than attaching data alongside the file, and is invisible to a person looking at it.

Does SynthID survive a screenshot?
Yes. That is the main practical difference between it and Content Credentials. Because the mark is in the pixels rather than attached metadata, screenshotting, cropping, compressing and format conversion all leave it detectable. Metadata does not survive any of those.

Can I check whether an image has SynthID?
Google operates a detector for it, and has been moving verification closer to where people look, including into Search. What you cannot do is verify an arbitrary image with an offline tool, because detection requires the matching detector rather than a published algorithm anyone can run.

Does no SynthID mean an image is real?
No, and this is the most misunderstood part. Most generated content comes from tools that mark nothing. An unmarked file might be a real photograph, might be from an unmarked generator, or might be from a marked one that was never covered. Presence is evidence; absence is silence.

How is SynthID different from C2PA Content Credentials?
C2PA attaches a cryptographically signed record to the file describing who made it and what was edited. SynthID changes the content itself. C2PA carries far more detail and dies to a screenshot; SynthID carries only the mark and survives one. They cover each other's weakness, which is why providers increasingly use both.

Is SynthID watermarking reliable for text?
Much less so than for images. It works by biasing word choices during generation, so a short passage carries too little signal to detect, and paraphrasing degrades it substantially. Treat text watermarking as a weak signal rather than a determination.