Every day, more AI-generated images, songs, videos, and articles enter the feeds we read and the sources we cite. That raises a practical question for anyone who works with content: was this made by a person or a model? Google's answer is SynthID, an invisible watermark that Google DeepMind embeds into content produced by its AI systems.
SynthID is one of the most ambitious provenance efforts shipping today. By May 2026, Google reported that more than 10 billion pieces of content had been watermarked with it. That is a serious number, and it is worth understanding what SynthID actually does, where it helps, and, just as important, where it does not.
This guide explains SynthID in plain English for journalists, content teams, and curious readers. No hype, and no pretending it is something it is not.
What SynthID is
SynthID is a family of watermarking tools built by Google DeepMind. The goal is simple to state: mark content that Google's AI models generate so that the content can later be identified as AI-made, even after it has been shared, cropped, or re-encoded.
The key word is invisible. A SynthID watermark is not a visible logo in the corner of an image or a tag buried in a file's metadata. Metadata gets stripped the moment you screenshot an image or paste text into a new document. SynthID instead lives inside the content itself, woven into the pixels, the waveform, or the word choices so that ordinary editing does not automatically remove it.
There are now separate versions for different kinds of content. SynthID first arrived for images in 2023, expanded to text and audio, and by 2026 covers images, audio, video, and text. Each version works differently under the hood, because a pixel grid and a sentence are very different things.
How invisible watermarking works, across image, audio, video, and text
The core idea is the same across all four media types. During generation, the model makes thousands of tiny choices, and SynthID nudges those choices in a pattern that a human cannot perceive but a matching detector can recognize. Here is how that plays out for each type.
Images. An image is a grid of pixels, and there is a lot of room to make micro-adjustments that the eye will never notice. SynthID embeds a pattern into the pixel data at the moment the image is created. Because the signal is spread across the whole image rather than sitting in one corner, it is designed to survive common edits like cropping, applying filters, compression, or resizing.
Audio. For sound, SynthID converts the audio into a spectrogram, a visual representation of the frequencies over time, then adds an inaudible signal and converts it back to a waveform. You hear normal music or speech. The watermark rides along in parts of the signal your ears do not register, and it is built to survive changes like added noise or speed adjustments.
Video. Video is a sequence of image frames, so SynthID watermarks the frames as they are generated. The same resilience goals apply, so that re-encoding or compressing a clip does not wipe the signal.
Text. Text is the hardest case, because you cannot slightly alter a word the way you can slightly alter a pixel. A word is either there or it is not. So SynthID-Text works during generation. As the model picks each word, it is choosing from a list of probable next words. SynthID gently adjusts those probability scores using a pseudorandom pattern, so the model leans toward certain word choices in a way that forms a detectable statistical signature. Done carefully, this does not degrade the quality, accuracy, or speed of the writing. Google has open-sourced the SynthID-Text approach, and a production-grade implementation ships in the Hugging Face Transformers library, so other developers can adopt it too.
Which Google products use SynthID
By 2026, SynthID is embedded across Google's generative products, and the watermark is applied automatically when content is created. The main integrations are:
- Gemini for text output
- Imagen for generated images
- Veo for generated video
- Lyria for generated music and audio, along with audio from tools like NotebookLM
In practice, if you generate an image in Imagen or a track in Lyria, the SynthID watermark is added by default. You do not switch it on, and typically you cannot switch it off. That default-on behavior is what makes the 10-billion figure possible.
How detection of the watermark works
A watermark is only useful if you can read it back. SynthID detection works by looking for the specific pattern that was embedded, a matched-filter approach where the detector knows what signature to search for.
Google runs a public SynthID Detector portal. You upload an image, audio, video, or text, and the portal scans it for a SynthID watermark. Rather than giving only a plain yes or no, it can highlight which parts of the content are most likely watermarked, for example pinpointing segments of an audio clip or areas of an image. You can also upload files directly to Gemini to check them.
It is worth being precise about what the result means. Detection is probabilistic, not a certificate. For text in particular, the detector returns one of three answers: watermarked, not watermarked, or uncertain, and you can tune the thresholds that trade off false positives against false negatives. When the portal launched, video and text detection were still rolling out and initial access was limited to journalists and media professionals, so exact availability keeps changing.
The honest limits of SynthID
This is the part that matters most, and it is where a lot of coverage oversells the technology. SynthID is a genuinely useful provenance tool, but it is not a universal AI detector. Three limits are essential to understand.
1. It only flags content from participating models. SynthID can find its own watermark, and nothing else. If content was generated by a model that does not embed SynthID, there is no watermark to find, so the detector will simply come back negative. That negative does not mean the content is human-made. Most AI content circulating online today comes from models outside Google's ecosystem, including many open-source image and text generators, and none of it carries a SynthID mark. A clean SynthID result tells you "no Google watermark here," not "a human wrote or drew this."
2. Heavy editing can weaken or remove the signal. SynthID is engineered to survive routine changes, and that resilience is real. But it is not indestructible. For text, Google is direct about this: detector confidence can be greatly reduced when AI-generated text is thoroughly rewritten or translated into another language. The same principle applies across media. Aggressive editing, re-generation, or format changes can degrade the watermark to the point where it no longer reads. Text also has a structural weak spot, because factual, short, or highly constrained writing leaves little room to embed a signal without changing the words, so those cases are harder to watermark reliably.
3. It is not designed to answer "is this AI?" in general. SynthID answers a narrower question: "did a participating Google model make this and leave its mark?" That is a provenance question, not a detection question. Provenance and detection are different jobs, and conflating them leads to false confidence in exactly the situations where you most need to be careful.
Why statistical AI detection is still necessary
Here is the practical takeaway. Because the vast majority of AI content in the wild is not SynthID-watermarked, a watermark check alone will miss most of it. Watermarking is a strong signal when it is present, and it is close to useless when it is absent, which is most of the time.
That is why statistical detection remains a necessary complement. Instead of looking for a planted signal, statistical detectors analyze the content itself for the patterns that machine generation tends to leave behind, and they work regardless of which model produced the content or whether anyone chose to watermark it. If you are verifying an image whose source you do not know, an image detector can assess it even with no watermark present. The same logic applies to audio with a voice detector and to writing with an AI detector.
To be equally honest about our own side of the fence: statistical detection has real limits too. It produces probabilities, not proof, and it can be wrong in both directions. We lay those out plainly on our page about AI detection limitations. The responsible approach is to treat SynthID and statistical detection as two imperfect tools that cover for each other's blind spots, and to treat any single result as evidence rather than a verdict.
FAQ
Does a negative SynthID result mean the content is human-made? No. A negative result means no SynthID watermark was found. The content could still be AI-generated by a model that does not use SynthID, which describes most AI content online. Treat a negative as inconclusive, not as proof of human origin.
Can SynthID watermarks be removed? They are built to survive routine edits like cropping, compression, filters, and format changes. But heavy editing, thorough rewriting, or translation can weaken or remove the signal, especially for text. So watermark absence never rules out AI generation.
Which content does SynthID cover? It covers images, audio, video, and text produced by participating Google models, mainly Imagen for images, Veo for video, Lyria for audio, and Gemini for text. Content from other companies' models is not watermarked by SynthID.
Should I rely on SynthID alone to verify content? No. SynthID is excellent at confirming that a participating Google model made something, but it cannot flag AI content from other sources or content whose watermark has been stripped. Pair it with a statistical detector for anything whose origin you cannot confirm.
Try it on your own writing