What Is SynthID? How Google's AI Watermark Works, How to Check It and Its Limits (2026)

Is that image or video you saw on social media a real photo, or was it made with AI? As it becomes harder to tell just by looking, technologies that place an "invisible mark" in AI-generated content so that machines can check it later are spreading. A leading example is Google DeepMind's SynthID. Based on Google's official blog posts and help pages, this guide walks through how SynthID works, which services use it, how users can check for it and its limitations, as of October 9, 2026.

What SynthID is

SynthID is a Google DeepMind technology for adding a digital watermark to content generated or edited with AI and detecting it later. It covers four types of content: images, video, audio and text.

This is not a visible watermark like a logo laid over the corner of an image. The signal is embedded in the content itself (the pixels of an image, the waveform of audio) in a way that human eyes and ears cannot notice. As a result, the watermark stays with the content even if the file's attached information (metadata) is removed.

According to Google, more than 180 billion images and videos and the equivalent of 240,000 years of audio have been watermarked with SynthID so far (announced October 7, 2026).

Timeline

August 2023: released in beta for some users of the image generation model Imagen on Google Cloud's Vertex AI
October 2024: SynthID Text for text released as open source, with the method published as a paper in the scientific journal Nature
May 2025: SynthID Detector, a portal for checking watermarks, announced (initially through a waitlist for journalists and researchers)
November to December 2025: the Gemini app becomes able to check whether images and videos contain SynthID
May 2026: at Google I/O 2026, verification in Search and Chrome is announced, along with adoption by OpenAI, Kakao and ElevenLabs
October 7, 2026: SynthID Detector opened to everyone (in English, available worldwide)

How it works: adding an invisible watermark

SynthID combines a model that "embeds" the watermark with a model that "detects" it. The characteristics for each media type are as follows.

Images and video: the watermark is embedded in the pixels without changing image quality. It is designed to survive edits such as cropping, filters, frame rate changes and compression.
Audio: the watermark is embedded in a way the human ear cannot hear. Google says it withstands added noise, MP3 compression and changes to playback speed.
Text: the probabilities the AI uses to choose the next word are adjusted slightly, leaving the watermark as a pattern in word choice. Google says this does not affect the quality of the text.

Because the mechanism for text differs greatly from images and video, it is covered in detail in "What is SynthID Text?".

Services that use SynthID

According to Google's official information, the main services whose output carries a SynthID watermark are:

Images and videos generated in Gemini (app and web)
Output from Imagen (image generation), Veo (video generation) and Lyria (music generation)
NotebookLM Audio Overviews
Images edited with "Reimagine" in Google Photos Magic Editor

Outside Google, NVIDIA adds SynthID to the output of its video generation model Cosmos. In May 2026, OpenAI, Kakao and ElevenLabs announced that they would adopt SynthID. Google also says content generated by Apple will soon be checkable in SynthID Detector.

How to check whether content has SynthID

There are three main ways for users to check for a SynthID watermark.

Checking in the Gemini app

Upload an image, video or audio file to the Gemini app and ask something like "Was this made with AI?", and it will check for SynthID and tell you. For video, it even shows in which segments of the visuals and audio the watermark was found. However, for now the Gemini app can only identify content made with Google's AI. According to the help page, the main conditions are:

One file at a time, up to 100MB. Video must be under 90 seconds and audio under one hour
Up to about 10 checks per day each for images, video and audio (video up to 5 minutes in total, audio up to 3 hours in total)
You must be signed in with a personal Google Account or an eligible Workspace account

Checking with SynthID Detector

Since October 7, 2026, SynthID Detector (synthid.com) has been available to everyone. Upload an image, video or audio file to check whether it contains a watermark. It is available in English and also supports content created with AI from OpenAI, NVIDIA and Kakao in addition to Google.

Checking in Search and Chrome

At Google I/O in May 2026, Google announced the ability to ask "Was this made with AI?" from Google Lens, AI Mode in Search, Circle to Search and Gemini in Chrome.

Limitations and caveats of SynthID

SynthID is useful, but it is not all-powerful. Google itself states limitations such as the following.

No watermark does not mean authentic: SynthID can only identify content made with AI from Google and the companies that have adopted it. Even if no watermark is found, the content may still have been made with another company's AI.
Not perfect against extreme edits: if an image is heavily reworked, the watermark may no longer be detectable.
Small edits may not be covered: Google Photos says that edits as minor as changing the color of a small flower may not be watermarked.
Text has particular weaknesses: detection accuracy drops for short factual answers, heavily rewritten text and text translated into another language.

It is safest to treat a watermark result as one clue for judging authenticity, and to combine it with checking the source and other information.

How it differs from C2PA

Besides digital watermarks like SynthID, another technology for showing where AI content came from is C2PA (Content Credentials). C2PA attaches provenance information to a file as signed data, such as when it was created, with which tool and how it was edited.

SynthID: embedded in the content itself, so it tends to survive even when metadata is removed. However, it only tells you whether the content was made with a supporting AI
C2PA: can record the history of creation and editing. However, the information can be lost through screenshots or re-saving

Google's approach is to combine the two, and it also attaches C2PA information to images created in the Gemini app and elsewhere. C2PA is covered in "What is C2PA?", and the differences between methods are explained in detail in "AI watermarks compared".

Watermarking at other AI companies

Efforts to mark AI output are spreading beyond Google.

OpenAI: announced its adoption of SynthID in May 2026.
Microsoft: attaches C2PA provenance information to images generated with Azure OpenAI.
Meta: has released Video Seal, a watermarking technology for video, as open source.
Anthropic: adds an invisible watermark to text generated by Claude. Models released on or after August 2, 2026 support it from launch, and earlier models are also watermarked on Anthropic's own services. Newer models also attach C2PA information to the files they generate. According to Anthropic, this text watermark is based on the SynthID Text method. Watermark detection is currently a private preview limited to regulators and similar bodies.

Regulation is also part of the background. The EU AI Act requires providers of generative AI to make their output detectable by machines as AI-generated. This provision has applied since August 2, 2026. However, under the 2026 amendment to the AI Act (the Digital Omnibus), only AI systems placed on the market before that date get a grace period until December 2, 2026 for the machine-readable marking obligation.

Summary

SynthID is a Google DeepMind technology that adds an invisible digital watermark to AI-generated images, video, audio and text so it can be checked later. In October 2026, SynthID Detector was opened to the public, letting anyone check the files they have. Adoption is also spreading to other companies such as OpenAI.

However, not finding a watermark does not prove that a person made something. Use SynthID and C2PA results as one piece of evidence, and look at them together with checking the source.