What Is SynthID? Google’s AI Watermark for Text, Images, Audio, and Video

SynthID is the most widely deployed AI watermark in existence and the most misunderstood, mostly because one name covers four different systems. Here is what each one is, where it is used, and what anyone outside Google can do with it as of September 14, 2026.

Four systems, one name

Google DeepMind introduced SynthID in 2023 as an invisible watermark for images from Imagen, extended it to audio later that year, and added text and video in 2024. The text version was published in Nature in October 2024 and open-sourced the same week. The four share a name and a purpose. They do not share a mechanism.

• SynthID Text is a keyed bias in word choice.
• SynthID Image is a signal embedded in the pixel values.
• SynthID Audio is a signal embedded in the waveform.
• SynthID Video marks each frame the way the image system does.

When a page says "SynthID detector", the first question is which of the four it means. The answers differ for every one.

How the text watermark works

Every time a language model picks the next word, it has several candidates it considers roughly equally good. SynthID Text uses a secret key to score those candidates with a pseudorandom function and tilts the choice toward the higher-scoring ones. A reader cannot see the tilt. A party holding the key can measure it across a whole passage and say whether the pattern is there.

Three consequences follow. Nothing is added to the text: no hidden characters, no metadata, no formatting. Longer text scores more reliably, and short or highly factual text, where the model has little freedom in word choice, scores weakly or not at all. And because the signal is in the words, a thorough rewrite or a translation through another model removes it, while light edits only weaken it.

Anthropic credits SynthID Text as the basis for Claude's watermark, which is why the two behave alike and why neither can read the other's key.

How the image watermark works

SynthID Image changes pixel values across the picture in a pattern that a matched detector can recognise and a viewer cannot. It is not a logo, not a metadata field, and not the sparkle Google draws on some Gemini images. It is designed to survive the transformations people assume defeat a watermark: screenshots, JPEG compression, resizing, cropping, colour changes and filters.

Google usually ships C2PA Content Credentials in the same file. That is a separate signal, a signed manifest in the metadata, and it does not survive a screenshot. The Gemini watermark page keeps the three marks apart.

Who uses it

Google says essentially all Gemini text carries SynthID Text, across the app and the API, and an independent test on August 23, 2026 found the signal in Gemini 3.5 Flash output. Images from Gemini and Imagen, audio from Lyria and video from Veo carry their respective marks.

Google also licenses it. Since May 19, 2026, every image from ChatGPT, the OpenAI API and Codex carries SynthID in the pixels, and OpenAI audio has carried it since July 31, 2026. That licensing deal produced the one public SynthID check that exists today, which is OpenAI's, and it reads OpenAI's marks only.

What SynthID is not

It is not a classifier. It does not guess from writing style or image texture; it reads a signal that was deliberately put there, with the key that put it there. That is why a SynthID verdict is a different kind of statement from a GPTZero score, and why the vocabulary matters.

It carries no identity. A positive result says a Google model produced or processed the content. It does not say which account, which prompt, or when.

It is not Content Credentials. C2PA is a label in the file; SynthID is a mark in the content. One can be stripped in a metadata pass. The other cannot.

Where anyone can check it

This is the part that decides what SynthID means in practice, and it is uneven.

• Text: Google's SynthID Detector portal, announced in 2025, is waitlist-only for journalists, media professionals and researchers. There is no public paste-and-check page. The open-source SynthID-Text repository publishes the code but not Google's production key, so it can score text you watermarked yourself and nothing else.
• Images, video and audio from Google: verification is live inside Google's own surfaces. Upload the image, video or audio clip to the Gemini app and ask whether Google AI created or altered it. For images, Lens, AI Mode and Circle to Search in Google Search run the same check, and Chrome is rolling it out. A limited-preview Cloud API exists for partners.
• Images and audio from OpenAI: OpenAI's Verify tool and Content Provenance API check the SynthID OpenAI embeds. DetectMark runs that check on one consent click through the OpenAI image verifier.
• Everything else: no public route. Any website offering a SynthID check outside those surfaces is either reading Content Credentials, which is a different signal, or running a classifier.

The SynthID detector page keeps this list current, row by row and dated. How to check for SynthID turns it into steps.

Why it matters more than it used to

The EU AI Act's Article 50 became enforceable on August 2, 2026 and requires generative AI output to be marked in a machine-readable, detectable way. SynthID is the most mature scheme that meets that description, which is why Google's competitors have licensed it or built on it rather than starting over. For anyone checking content, the consequence is simple: the marks are increasingly there, and the ability to read them is still mostly not.