What are Google's SynthID watermarks and how do they help users tell real content from AI-generated content

A particular kind of uncertainty has crept into how people interact with online content. A photograph of a public figure making a statement, a video of an event that may or may not have happened, or an audio clip that sounds completely authentic—none of these can be taken at face value the way they once could. Generative artificial intelligence (AI) has made it possible to produce content that is indistinguishable from something a human captured or recorded to the eye and ear.


The problem is not that AI-generated content exists. It is that there has historically been no reliable way to tell what is real from what has been synthesised, short of technical forensic analysis that most people are not equipped to carry out. The tools for identifying altered images, detecting inconsistencies in shadows, checking metadata, and running reverse image searches require effort, knowledge, and time. And even then, they can fail.

Google's response to this problem is SynthID, a watermarking technology developed by Google DeepMind that embeds invisible signals directly into AI-generated content. The watermark is not a logo or a banner. It is not something that can be cropped out or edited away with any ease. It exists within the content itself, within the pixels of an image, the frequency structure of an audio file, the frames of a video, and the statistical patterns of generated text.

At Google I/O 2026, Google announced a significant expansion of where and how SynthID works, making it one of the more consequential announcements from an event already full of them. Understanding what SynthID actually is, how it works, and where it is now available requires untangling a few threads that often get tangled in coverage of AI safety tools.

Why traditional watermarks were not adequate for this problem

Watermarks have existed in one form or another for centuries. Physical paper watermarks are visible when held to the light. Translucent logos placed over photographs. Metadata embedded in digital files that records when and where an image was taken and with what device. Each of these serves a legitimate purpose, but none solves the specific problem posed by AI-generated content.


The issue with visible watermarks is obvious: they can be removed. Cropping an image eliminates a corner watermark. Image editing tools can remove translucent overlays. Generative models themselves can sometimes be prompted to produce content without the visible branding a platform might add.

Metadata is a different matter, but it comes with its own fragility. When a file is shared across platforms, re-uploaded, screenshotted, or converted between formats, metadata is often stripped away entirely. A photograph taken with a camera that embeds GPS coordinates and timestamps in the file loses all that information the moment someone takes a screenshot and shares it. The chain of provenance breaks, and there is no way to reconstruct it.

SynthID approaches the problem differently. Rather than attaching identification information to a file in a way that can be separated from it, SynthID embeds the watermark into the content itself. For images, this means altering pixel values in ways that are statistically detectable but visually imperceptible.

The watermark is not a sticker on top of the image; it is woven into the image. Modifications such as changing colour balance, adding filters, applying lossy compression, or resizing the image do not necessarily remove the watermark, since it is distributed throughout the content rather than localised in a single removable spot.

How SynthID works across different content types

The first version of SynthID was created for image data, and the beta version was released in 2023, when the system went live on Google Cloud as a companion to Google’s text-to-image AI model, Imagen. In the first iteration, two models were jointly trained using machine learning. Their union was a compromise of conflicting interests.

The watermark needed to be invisible enough that it did not affect how the image looked or its usefulness for creative purposes. And it needed to be detectable enough that the identification model could reliably find it even after the image had been through normal transformations.

Since that initial launch, SynthID has expanded considerably. It now covers audio content generated by Lyria, Google's AI music model. It covers video content from Veo. And it covers text generated by large language models, through a technique called SynthID Text, which was open-sourced to allow developers outside of Google to build the same capability into their own models.

Text watermarking differs from image or audio watermarking because text lacks pixel values or frequency components to modify. Instead, SynthID Text works at the level of token selection during generation. When a language model is generating text, it selects each token, which is roughly every word or word fragment, from a probability distribution.

SynthID Text uses, as described in the technical documentation, a pseudorandom g-function applied as a logits processor to subtly influence which tokens are selected during generation. The influence is small enough that it does not meaningfully change the quality or meaning of the output. But the pattern of choices it creates is detectable by a trained classifier.

The detection side of SynthID Text is probabilistic. It produces three possible outcomes: watermarked, not watermarked, or uncertain. The threshold between these outcomes can be tuned depending on how much risk of false positives or false negatives a given application can tolerate.

A useful limitation to consider is that the watermarks produced by SynthID Text are not resistant to all forms of modifications. According to their technical details, these watermarks remain immune to cropping text, changing some words, and minor paraphrasing; however, full rewriting or translating the text into any other language will severely hamper the system's ability to detect the watermark with high confidence.

What SynthID Detector is and who it is for

Last year, Google introduced a portal called SynthID Detector, which allows users to upload content to check for a SynthID watermark. The portal supports images, audio, video and text. The return if a watermark is found is more than a positive result. For images, it shows the parts most likely to contain a watermark. For audio, it searches for specific segments where the watermark is present.


The SynthID Detector is starting to ship to early testers, with a waitlist prioritising journalists, media professionals and researchers. That is just the reality of who this kind of tool will benefit most in the near term. A journalist verifying whether a piece of video footage is authentic before publishing a story, a fact-checker assessing a viral image, a platform trust-and-safety team evaluating content at scale, these are the people for whom a reliable, dedicated verification portal matters most.

The portal remains in a somewhat limited access phase, but the broader ecosystem around it has been expanding. Google open-sourced SynthID Text watermarking to allow other developers and model providers to embed compatible watermarks in their own outputs. A partnership with NVIDIA means that video generated by Nvidia's Cosmos NIM microservice also carries SynthID watermarks. And a partnership with GetReal Security, described in Google's announcements as a content verification platform, extends detection capabilities further.

The logic behind these partnerships reveals an important aspect of how content travels online. A watermarking system that covers only content generated by one company's models is, by definition, limited. If the goal is to create a reliable way for people to identify AI-generated content, the watermark needs to be present in content from as many sources as possible, and the detection needs to work regardless of where the content was found or how it was shared.

What Google I/O 2026 changed

At Google I/O 2026, the SynthID ecosystem expanded in several directions simultaneously. The most immediately relevant for ordinary users is that SynthID verification moved beyond the Gemini app and the dedicated SynthID Detector portal and into Google Search and Chrome.

According to Google's announcements from the event, users can now check whether an image was generated by AI using Search features, including Lens, AI Mode, and Circle to Search, as well as Gemini in Chrome. The interaction is designed to be conversational, asking something like "Is this made with AI?" or "Is this AI-generated?" rather than requiring the user to navigate to a separate tool or understand anything about how watermark detection works.

This matters because it removes friction from the verification process. Previously, someone who came across a suspicious image online would have needed to save the image, navigate to a separate tool, upload it, and wait for a result. That is a sequence of steps that most people simply will not complete, particularly in the middle of reading an article or scrolling through a feed. Bringing the check directly into Search and Chrome makes it something that can happen in the moment, at the point where the content is actually being encountered.

Along with SynthID, Google also announced a wider rollout of Content Credentials, a different, but related, technology that works more like a digital passport for content. SynthID is intended to detect AI-generated or AI-modified content, whereas Content Credentials is a specification for tracking the source and provenance of any content, whether or not it was created with AI. The two technologies serve different goals, but Google is increasingly using them together. The Gemini app got Content Credentials verification, with rollout kicking off around I/O 2026, and support for Search and Chrome in the coming months.

The numbers Google cited at I/O 2026 give a sense of the scale at which SynthID is now operating. The Gemini app's SynthID verification feature, added before the I/O announcement, had already been used 50 million times globally by the time the announcement was made. And since the technology's launch, more than 100 billion images, videos and content representing 60,000 years of audio have been watermarked with SynthID across Google's generative media tools.

The limits of what SynthID can do

No discussion of SynthID would be honest without acknowledging what it cannot do. Google's own documentation on SynthID Text notes that the technology "is not designed to directly stop motivated adversaries from causing harm." A person determined to use AI-generated content to mislead others has several options that SynthID does not close off. They can use AI models that do not embed SynthID watermarks. They could subject their output to enough human editing that the watermark would be degraded. They could also deploy AI models that do not fall under the SynthID umbrella.

This defence against the accusation, found in Google's documentation, is that SynthID "could help to hinder any malicious usage of AI-generated content," since it could be used "alongside other tools to cover all angles." The statement is somewhat more balanced than those above, and it’s also reasonable to believe.

A lock isn’t an absolute barrier; it increases the cost and hassle. A watermark won’t stop the spread of false information generated by AIs; however, it will make spreading that information harder when the watermark is present on that AI-generated content.

The expansion of the SynthID ecosystem to include other AI companies, such as OpenAI, Kakao, and ElevenLabs, was announced at I/O 2026, bringing SynthID technology into their outputs and narrowing the gap between what is covered and what is not. The more models that embed compatible watermarks, the more useful the detection capability becomes for ordinary users who have no way of knowing which AI tool produced a given piece of content.

How users can access SynthID verification today

For someone who wants to use SynthID verification now, the options are more varied than they were even six months ago. The Gemini app supports uploading images, videos, and audio files, and allows users to ask whether the content was created or edited by Google AI. The same capability is available through the Gemini web interface at gemini.google.com.

Through Google Search, Lens, and Circle to Search, verification can happen directly from within search results or from content seen on a mobile screen. Chrome is receiving Gemini-powered verification as part of a rollout that was announced at I/O 2026.

For journalists and media professionals who need a more detailed view of where a watermark is located within a piece of content, the SynthID Detector portal is available to early testers via a waitlist.

The technical requirements are modest. A Google account, the latest version of the Gemini app if using it on mobile, and a signed-in session. Files need to be under 100 MB, videos under 90 seconds, and audio files under an hour.

What do these points point toward

The problem SynthID is trying to solve is not one that any single technology will fully resolve. Content authenticity is a challenge that involves technical systems, platform policies, media literacy, regulatory frameworks, and the incentives that shape how content gets created and shared. SynthID addresses one part of the technical capacity to embed and detect a signal in AI-generated content and does so in a way designed to remain functional even as content is shared, modified, and moved across platforms.


What has changed at I/O 2026 is not the underlying technology but its reach. Verification is no longer something a user has to specifically seek out. It is becoming part of the tools people already use to find and interact with information. That shift in accessibility is, in practical terms, what determines whether a verification technology gets used or sits unused on a dedicated page that most people never visit.


The scale of content being watermarked over 100 billion pieces and the pace at which other AI providers are joining the ecosystem suggest that the infrastructure for content verification is being built out with some seriousness. Whether that infrastructure ultimately proves adequate to the scale of AI-generated content in circulation remains genuinely open. But the direction of travel is clear enough, and for anyone trying to understand what SynthID is and what role it plays, that direction is worth paying attention to.