Since 7 October 2026, Google’s SynthID Detector has been publicly available worldwide. At synthid.com, users can upload images, videos and audio files to check whether they contain an invisible SynthID watermark. The interface is initially available only in English. Before this public launch, the portal had been available since 2025 mainly to journalists, media professionals and selected researchers.
The move matters because determining the origin of digital content is becoming increasingly difficult. AI-generated images, voices and videos are growing more realistic, traditional visual clues are becoming less reliable and synthetic media is spreading rapidly across social networks, messaging services and corporate communications. Google’s approach with SynthID is therefore different: rather than trying to guess afterwards whether something looks AI-generated, the system embeds a machine-readable signal into the content at the point of creation.
The public Detector now makes that signal easy for anyone to check.
SynthID Does Not Look for AI Characteristics – It Looks for a Watermark
The most important thing to understand is that the SynthID Detector is not a conventional AI detector.
Many so-called AI detectors attempt to estimate whether content was probably generated by AI by looking for statistical anomalies, visual mistakes, linguistic patterns or other tell-tale signs. Such methods inevitably produce false positives and can become unreliable very quickly as models improve.
SynthID takes a different approach.
Google DeepMind embeds an invisible digital watermark directly into the content during generation. In images and videos, the signal is encoded into the visual information; in audio, it is embedded in the sound signal. The watermark is designed to be imperceptible to people and not to visibly or audibly reduce the quality of the content. The Detector then searches specifically for this embedded pattern.
That makes a positive result comparatively meaningful. If the Detector finds a valid SynthID signal, it indicates that at least part of the content was created or edited using a compatible AI system.
A negative result means something very different.
A Negative Result Does Not Prove That Content Is Authentic
This is the most important limitation of the entire system.
If the SynthID Detector finds no watermark, that does not mean the image, video or audio file was created by a human. It means only that no detectable SynthID watermark is present.
The content could still have been generated by a different AI system. It may have come from a model that does not support SynthID. It may have been created before the watermark was introduced. Or the signal may have been weakened by heavy editing, repeated conversion or other transformations to the point where it can no longer be detected.
Google itself makes clear that SynthID is not a general authenticity check. In Gemini as well, the absence of a watermark does not mean that a piece of media was not created using artificial intelligence.
For journalistic verification or content moderation, the practical rule is simple: a positive result is a relevant indication of AI involvement. A negative result is not proof of authenticity.
Google Is Opening SynthID to Other Providers
One of the most important changes with the public version is that the Detector is no longer limited to Google products.
Google now names OpenAI, NVIDIA and Kakao as partners whose AI-generated media can be watermarked with SynthID and detected through the portal. Apple is also expected to join the system.
This shows that Google is trying to turn SynthID from an internal labelling technology into a broader industry standard.
That is strategically important. A watermark is only truly useful if as many generators as possible use the same or compatible technology. If every provider relied exclusively on its own proprietary standard, users could end up needing a separate verification tool for each platform.
That problem has not been solved yet. Other companies use different provenance and watermarking systems. At the same time, C2PA and Content Credentials are gaining ground as a separate approach for carrying information about provenance and editing history.
SynthID is therefore best understood as one part of an emerging verification ecosystem rather than as the single universal standard.
Google Has Already Watermarked Enormous Volumes of AI Media
The scale of the system is now substantial.
According to Google, more than 180 billion images and videos have been watermarked with SynthID since the technology launched in 2023. In addition, audio carrying the watermark now represents a combined duration of around 240,000 years. Verification features integrated into Google Search, Gemini and Chrome reportedly handle more than one million checks per day.
These are Google’s own figures and should therefore be treated as company-reported numbers. Even so, they show that SynthID is no longer a small research experiment.
Google already integrates the technology across its own image, video and audio models, as well as across Gemini products. DeepMind now describes SynthID as a cross-media watermarking technology for images, video, audio and text.
The User Experience Is Deliberately Simple
No specialist software is required to use the public SynthID Detector.
After visiting synthid.com, users first need to sign in. At present, Google, Apple and ChatGPT accounts are supported. A file can then be uploaded or dragged into the browser for analysis.
The Detector checks whether a SynthID watermark is present. For images, the result is essentially a yes-or-no indication of whether a watermark has been detected. For video and audio, the system can additionally indicate the sections in which the signal was found.
Early tests by technology publications suggest that analysis typically takes only a few seconds.
A Wide Range of Image, Video and Audio Formats Is Supported
The public Detector currently accepts a fairly broad selection of common file formats.
For images, supported formats include JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF and GIF. For video, the supported formats include MP4, MOV and WEBM. For audio, users can upload WAV, MP3, OGG, FLAC, AAC and M4A files.
That covers most of the formats commonly used across the web, smartphones and social media.
Google has not published equally clear size and duration limits for the standalone SynthID Detector as it has for Gemini. However, independent tests suggest that there are usage quotas of roughly ten checks per day for each media type. These limits appear to be part of a quota system and should not be treated as a permanently guaranteed technical restriction.
Verification in Gemini Works Slightly Differently
Users who do not want to use the standalone Detector can also perform SynthID checks within the Gemini app.
There, an image, video or audio file can be uploaded and the user can ask, for example, whether it was created or edited using AI.
Gemini has clearly defined limits. A file can be no larger than 100 MB, videos must be shorter than 90 seconds and audio files must be shorter than one hour.
There is, however, an important difference compared with the standalone Detector.
Google states that Gemini’s SynthID checks can currently identify only content created using Google AI tools. The broader partner support involving OpenAI, NVIDIA and Kakao applies to the public SynthID Detector and should not automatically be assumed to work in Gemini as well.
That distinction is easy to miss.
Google Is Increasingly Combining SynthID with Content Credentials
Alongside SynthID, Google is expanding a second layer of verification: Content Credentials.
This is a provenance system based on the C2PA standard. In simple terms, it works like a digital passport for a file. It can contain information about how content was created, which editing steps were performed and whether generative AI was involved.
SynthID and Content Credentials solve different problems.
SynthID is embedded directly into the media itself and can therefore survive even when conventional metadata is removed. Content Credentials, by contrast, can carry richer provenance information and document an editing history.
Google is increasingly combining both approaches in its products. In Gemini, users can check supported media for both SynthID watermarks and Content Credentials.
In the long term, that combination may prove especially important: a harder-to-remove signal embedded in the media itself, combined with a more detailed provenance record.
Everyday Editing Is Supposed to Leave SynthID Intact
One of the main technical challenges for any invisible watermark is robustness.
Google says SynthID is designed to survive common transformations such as resizing, colour changes, compression and other routine edits.
That matters because digital content is rarely passed around unchanged. Social media platforms automatically compress images and video, messaging apps alter file formats and publishers routinely crop or resize material.
A watermark that disappeared after ordinary JPEG compression would have very limited practical value.
SynthID is not indestructible, however. Google acknowledges that after repeated or substantial changes, the signal may no longer be detectable. Independent testing also shows that digital watermarks can, in principle, be weakened or removed in some circumstances.
SynthID is therefore robust, but not tamper-proof.
The Detector Cannot Reliably Tell How Much AI Was Involved
Another important limitation concerns AI-assisted editing.
SynthID is not used only for media that has been generated entirely from scratch. A file can also carry a watermark if only part of it was edited using a compatible AI system.
A positive result therefore does not automatically mean that an entire image or video is synthetic.
For example, an AI tool might alter only part of the background, add an object or replace individual elements. The presence of SynthID would then indicate AI involvement, but not necessarily how extensive that involvement was.
That distinction matters particularly in journalism and verification. “AI-edited” and “fully AI-generated” are not the same thing.
Text Is Not Included in the Public Detector
Although SynthID also includes a technology for text, the new public portal does not accept text documents for verification.
SynthID Text works differently from watermarking for images, video and audio. During text generation, the probability distribution over possible tokens is subtly influenced in a way that creates a statistical pattern that can later be tested.
Google has made parts of this technology publicly available. The associated detector works probabilistically and can return results such as “watermarked”, “not watermarked” or “uncertain”.
The public synthid.com portal, however, is currently limited to images, video and audio.
Text Watermarking Remains a Difficult Problem
Text also illustrates why watermarking is not a universal solution for AI provenance.
An image or audio signal contains millions of data points in which an invisible marker can be distributed. Text consists of comparatively few symbols. If a passage is rewritten, translated, shortened or paraphrased, a statistical watermark can be damaged very quickly.
That is another reason why the new SynthID Detector should not be treated as a general-purpose detector for every form of generative AI.
Its strongest use case is currently clearly in media files.
Privacy Matters When Using a Public Detector
Uploading media for verification inevitably means sending that file to an external service.
That is particularly relevant for confidential corporate material, unpublished journalistic content, internal audio recordings or images containing personal data.
According to current information about the service, files are processed for analysis and then deleted. However, a digital fingerprint of the file may be stored temporarily in order to enforce usage limits. Secondary documentation of the public portal indicates a retention period of around 24 hours for this purpose.
For highly sensitive material, organisations should still assess whether external cloud-based verification is legally and operationally acceptable.
There Is Still No General Public Detector API
For professional workflows, one major question remains open: can the Detector be automated?
As of October 2026, there is no generally documented public API for the new SynthID Detector portal that would allow organisations to check large volumes of files automatically.
That means synthid.com cannot currently be integrated easily as a standard verification step into n8n, Make, CMS or digital asset management workflows.
SynthID technologies do exist in developer and open-source contexts, particularly for text, but that is not the same as having a freely available API for the new multimedia Detector.
For newsrooms, platforms and larger organisations, this remains a significant limitation. A tool becomes genuinely scalable only when checks can be performed automatically through interfaces.
For Newsrooms, SynthID Is Best Used as an Additional Piece of Evidence
For journalistic and editorial verification, the SynthID Detector can still be very useful.
Imagine a newsroom receives a striking image through social media. A SynthID match could provide a strong indication that a compatible AI system was involved in creating or editing the image.
A negative result, however, should never be treated as clearance.
Professional verification still needs to include other methods, such as reverse image search, metadata analysis, Content Credentials, the original source, temporal context, geolocation, image forensics and, where possible, direct confirmation from the creator.
SynthID does not replace those methods. It complements them.
The Real Breakthrough Would Be a Shared Industry Standard
The most interesting part of the public launch is therefore not the web form itself.
What matters is Google’s attempt to build a broader technical ecosystem. OpenAI, NVIDIA and Kakao are already named as partners, with Apple expected to follow. If more major providers adopt SynthID, the usefulness of the Detector will increase significantly.
At the same time, this approach will continue to compete with and complement other provenance technologies such as C2PA.
In the longer term, a multi-layered system could emerge: invisible watermarks for machine-readable detection, Content Credentials for documented provenance and platforms capable of evaluating both signals automatically.
That would be considerably more robust than the current situation, in which users often have to judge authenticity largely from the visible content itself.
SynthID Changes the Question from “Does This Look AI-Generated?” to “Is There a Verifiable Provenance Signal?”
That is the real significance of the new public Detector.
As image, video and audio models become more capable, it is becoming increasingly difficult to identify synthetic content reliably through human perception alone. Looking for unusual hands, shadows, voices or other supposed visual artefacts becomes less dependable with each generation of models.
SynthID starts from a different premise.
Rather than asking the viewer to guess whether content was artificially generated, the system that creates the content is expected to leave behind a machine-readable provenance signal from the outset.
The public SynthID Detector now makes that principle accessible to a broad audience for the first time.
It does not solve the problem of AI verification. A universal standard is still missing, not every provider is covered and no watermark can be guaranteed to survive every possible transformation.
But it does shift the direction of the debate.
The key question may increasingly become less:
“Can I tell that this image was generated by AI?”
And more:
“What verifiable provenance signals does the file itself provide?”
