A convincing image arrives in a group chat. It appears to show a real event, but the original source is missing. Can a tool establish whether artificial intelligence helped create it?
On October 7, 2026, Google announced that SynthID Detector was becoming available globally in English, following an earlier version aimed at media professionals. The standalone service checks images, video and audio for SynthID watermarks.
For readers, publishers and businesses, it offers another way to investigate media before sharing it. Understanding the scope of the check is essential: a watermark can provide evidence of AI involvement, while an unsuccessful search leaves the content’s origin unresolved.
What Google has opened to the public
In its announcement, Google says the expanded detector can identify watermarked media made using AI from Google and participating partners, including OpenAI, NVIDIA and Kakao. Apple was described as a forthcoming addition.
This does not mean every file created by those companies’ products is necessarily identifiable. A useful check depends on the relevant content carrying a supported watermark and that watermark remaining detectable.
The public entry point is SynthID Detector. Its image, video and audio coverage should also be distinguished from the wider SynthID technology, which includes a separate approach to watermarking generated text.
How an invisible watermark works
According to Google DeepMind, SynthID embeds an imperceptible signal directly into generated content. A compatible detector searches for that signal.
This provides a different kind of evidence from judging whether a picture looks suspicious. The check can look for an identifier embedded during generation, rather than relying on visual clues such as unusual hands, inconsistent reflections or distorted lettering.
The signal also differs from a visible logo placed in the corner of a picture. Removing a visible label and checking for an embedded watermark are separate matters.
Google says the technology is designed to withstand common changes, including cropping and compression. However, its help documentation acknowledges that repeated alterations can leave a watermark undetectable. Resilience should therefore be understood as a design goal with practical limits.
How to interpret the result
The most useful question is what evidence the tool found in the submitted file. Avoid turning that answer into a broader conclusion about the entire scene or story.
| Outcome | What it supports | What to check next |
|---|---|---|
| A supported watermark is detected | Evidence of AI generation or editing in the content covered by the result | Which parts were affected and how the publisher describes them |
| No supported watermark is detected | The check found no detectable supported signal | The original source, file history and other evidence of origin |
| The result is inconclusive | The tool cannot establish the watermark status with sufficient clarity | A better original file and additional verification methods |
This is an interpretation guide, rather than a reproduction of the service’s interface labels.
The US Office of Research Integrity makes the central limitation explicit: missing watermark or provenance indicators do not establish authenticity. Content produced with an unsupported AI system may carry neither.
Equally, evidence of an AI edit does not explain its significance. An illustrative background, a translated caption and a fabricated person make different editorial claims. The result needs to be read alongside the content and its context.
SynthID Detector and Gemini have different coverage
Google also provides media verification inside Gemini. Uploading a supported file and asking whether Google AI created or edited it can trigger a watermark check.
However, Google’s Gemini verification documentation currently says that Gemini’s SynthID recognition covers content created with Google AI tools, even though other companies have adopted the technology.
That distinction matters. The standalone detector’s announced partner coverage should not automatically be assumed to apply to the Gemini interface.
For a check through Gemini, the documentation specifies one file at a time, no larger than 100 MB, with videos under 90 seconds and audio under one hour. These are Gemini requirements; users should follow the standalone detector’s own upload requirements when using that service.
What Content Credentials add
A watermark searches for a signal embedded in the media. Content Credentials, based on the C2PA standard, provide signed information about its recorded origin and editing history.
The C2PA explainer describes a system for checking that provenance information is associated with an asset and has not been tampered with. It also stresses that provenance alone cannot determine whether what the content depicts is true.
Imagine a genuine photograph from several years ago republished with a caption claiming it was taken yesterday. A reliable creation history could help expose the mismatch, but establishing the claimed event still requires checking the date, location and source.
For a publisher, the practical benefit is having more evidence to inspect. A signed history and a watermark check answer different questions and can contribute to the same investigation.
A practical check before sharing a file
For everyday publishing, DemystIA recommends a short sequence:
- Find the best available original. Request the source file and the first publication, rather than relying only on a repost or screenshot.
- Run the relevant checks. Use the detector for supported watermarks and inspect Content Credentials where available.
- Record the scope of the answer. Keep the file checked, service used, date and result together.
- Check the claim attached to the media. Investigate the alleged event, date and location independently.
The same habit applies when evaluating AI answers that include sources: inspect what the evidence actually establishes before accepting the surrounding claim.
Why broader access matters
DemystIA’s assessment is that public access makes watermark verification easier to include in an ordinary publishing workflow. It gives small editorial teams and individual users access to a check previously introduced for media professionals.
Its usefulness will depend on adoption across generation tools, reliable detection after distribution, and clear explanations of the results. For now, the practical gain is a more accessible source of evidence about how a file was made—evidence that works best alongside a check of who published it and what they claim it shows.
Sources
- Google — Public expansion of SynthID Detector, October 7, 2026
- Google DeepMind — SynthID technology overview
- Google Help — Media verification in Gemini, coverage and limits
- C2PA — Content Credentials and provenance explainer
- US Office of Research Integrity — Content provenance and verification limits