What Is SynthID? Google’s AI Watermark Explained

What is SynthID?

SynthID is a watermarking technology from Google DeepMind, built to mark AI-generated content at the moment it is created and to recognise that mark afterwards. Google’s SynthID page, as we read it in October 2026, describes it as a watermarking tool designed specifically for AI-generated content.

Google DeepMind first announced SynthID on August 29, 2023, as a beta for images made with Imagen, Google’s text-to-image model, available to a limited number of Google Cloud Vertex AI customers. It has since been extended well beyond images.

Google applies SynthID to content from its own models. Its SynthID page also says the Detector portal can check content made with AI from partners, naming OpenAI, NVIDIA and Kakao, with Apple listed as coming soon.

The idea behind it is simple. Instead of guessing afterwards whether something was generated, the generator marks its own output at the source, and a matching detector reads the mark later. That turns a hard question, “does this look AI-made?”, into an easier one, “does this carry a known signal?”, but only for content from tools that apply the mark.

How does the SynthID watermark work?

SynthID works by changing the content in ways that are imperceptible to people but measurable by Google’s detection technology. The mark is part of the content itself, not a label stored next to it, and the method depends on the medium.

What is the SynthID Detector and how do you use it?

The SynthID Detector is Google’s verification portal for checking whether a file carries a SynthID watermark. Google announced it at Google I/O in May 2025, at first for early testers. Its SynthID page now says that, after testing with journalists and media professionals, access is being expanded to everyone.

You can also check from the Gemini app. In November 2025, Google’s post “How we’re bringing AI image verification to the Gemini app” explained that you upload an image and ask a question such as “Is this AI-generated?”. Gemini’s help page, as we read it in October 2026, covers images, video and audio, and on supported platforms it also reads C2PA Content Credentials.

1. 01Open the Gemini app or gemini.google.com and upload a single file.
2. 02Ask whether it was created with Google AI, or whether it is AI-generated.
3. 03Read the answer: Gemini reports whether it found a SynthID watermark and, where supported, what the Content Credentials say.
4. 04Keep the stated limits in mind: one file at a time, up to 100 MB, videos under 90 seconds and audio under one hour, with daily usage limits.

Google’s SynthID page adds that Gemini, Search and Chrome will check for a SynthID watermark and tell you if they find one. That is the closest thing to a built-in Google AI checker today, and it is a watermark check, not a general detector.

Is SynthID Text open source?

Yes. Google DeepMind open-sourced SynthID Text in October 2024, the same month it described the method in Nature in a paper titled “Scalable watermarking for identifying large language model outputs”.

According to Google’s developer documentation for SynthID Text, a production-grade implementation ships in Hugging Face Transformers from version 4.46.0, with a reference implementation on GitHub. Developers can apply the watermark to their own model’s output and use the provided detector to check text later.

Text is harder to watermark than an image. A picture has millions of pixels in which to hide a signal, while a sentence has only a handful of word choices, and many of them are fixed by grammar or by the facts. The signal therefore builds up over longer passages and is weakest in short or tightly constrained answers.

The same documentation is candid about the limits:

• The watermark is less effective on factual responses, where there is less room to vary word choice without hurting accuracy.
• Detector confidence can drop sharply when text is thoroughly rewritten or translated.
• It holds up to light edits, such as changing a few words or mild paraphrasing.
• Google says it is not designed to stop motivated adversaries on its own.
• Detection is probabilistic, with watermarked, not watermarked and uncertain outcomes.

So a SynthID Text check only applies to text produced by a model that applied the watermark. Text from a chatbot that does not use SynthID carries no signal at all; for that, a statistical AI text detector is the fallback, read as an indication rather than proof.

Watermark vs provenance vs detection: what is the difference?

A watermark is a hidden signal placed inside content, provenance is a signed record attached to the file, and detection is an estimate made from the content alone. All three try to answer “was this made with AI?”, and each one fails in its own way.

The three are complementary. A watermark and a signed manifest are strong when present, because they record what happened at creation. Detection is the fallback when neither is there, which is common for content shared on social media. Our guide on what C2PA and Content Credentials are explains the provenance side in detail.

What can’t SynthID tell you?

SynthID cannot tell you whether content made with other AI tools is AI-generated. Gemini’s help page puts it plainly: if no watermark is found, the content was not created or edited by Google AI, but other AI systems could still have made it.

• Coverage is limited to Google’s tools and the partners that apply SynthID.
• Some results are inconclusive, for example when content has too few details to carry a watermark or an edit is too small to leave one.
• Heavily rewritten or translated text loses much of the SynthID Text signal.
• In the Gemini app, collages are not supported, and Google advises cropping screenshots tightly around the image.
• A watermark says how content was made, not whether what it shows is true, fair or misleading.

In short, a positive SynthID result is strong evidence about origin. A negative one narrows the question but does not answer it. A real photo can still be used to mislead, and a generated image can be a harmless illustration.

There is also a coverage question that no single company can solve. A watermark only helps when the generator applies it and the checker can read it. Until most tools do both, watermark checks will sit alongside provenance records and detectors rather than replace them.

Is there a Google AI detector for any content?

No. When people search for a Google AI detector or a Google AI checker, what Google offers is a watermark check, not a general detector. It looks for SynthID and, in the Gemini app, Content Credentials. It is not designed to judge whether content from any generator is AI-made.

That design is a deliberate trade-off. A watermark check can be very confident when the mark is present, because it is looking for a known signal. A general detector has to estimate from patterns, which covers more content but with more uncertainty. Our guide on whether AI detectors are accurate explains that trade-off.

In practice, Google AI detection and independent detectors complement each other. Use Google’s tools when you suspect Google AI or a partner tool, and use a format-specific detector when you do not know where the content came from, which is the usual situation with content shared online.

Does AnyFact read SynthID watermarks?

No. AnyFact does not read SynthID watermarks. If you suspect content was made with Google AI, use Google’s SynthID Detector or the Gemini app for that check.

What AnyFact does is check provenance first. Every file is checked for C2PA Content Credentials and EXIF/XMP metadata, and if they name a known AI generator, the verdict is “probably AI” with the reason stated. Then an open-source detector scores the content, and Claude gives a second opinion when the score is borderline. The verdict comes in three levels: probably human, uncertain or probably AI.

That makes AnyFact useful in the cases a watermark check cannot cover, such as content from generators that apply no watermark. The Content Credentials checker reads the provenance layer, and the AI image detector adds an open-source classifier and lists visible anomalies such as hands, lettering, shadows, reflections and textures.

If you can run both kinds of check, do. A SynthID result and an AnyFact report answer different questions, and they are most useful side by side: one tells you whether a known watermark is present, the other what the provenance data and the content itself suggest. Every AnyFact verdict comes with its signals, a plain-language explanation, the limits of the method and a downloadable PDF report.

How to check if content was made with AI: a step-by-step workflow

Combine the three approaches in order, from the most certain to the least. Stop when you have enough evidence for the decision you need to make.

1. 01Find the original file, not a screenshot or a re-upload.
2. 02If you suspect Google AI or a SynthID partner, check for a watermark with the SynthID Detector or the Gemini app.
3. 03Read the C2PA Content Credentials and the file metadata.
4. 04Run the content through an AI detector suited to the format.
5. 05Trace the source: who posted it first, and is there corroborating evidence?
6. 06Weigh the results together and write down what each step found, including what found nothing.

The order matters because the first checks, when they find something, are the most decisive. A watermark or a signed manifest naming a generator settles the origin question in a way no detector score can. When both are missing, you are back to estimates and context, and the honest answer may be “unverified”.

For video and audio the same order applies, with extra clues in motion and sound. See how to spot a deepfake video for what to look and listen for, and what a deepfake is for the main types you may meet.

Frequently asked questions