Guides

SynthID Detector: Check AI Images, Video and Audio

Google has opened SynthID Detector to everyone. Learn how to check images, video and audio, and why an undetected watermark does not prove a file is real.

APIXO TeamOctober 9, 2026

Start creating

Adult female detective in a navy blazer examining image, video and audio cards with a magnifying glass

SynthID Detector: AI Media Checks and Their Limits | APIXO

AI-generated concept cover with contemporary English-market casual tailoring; illustrates watermark checking, not a real interface or detection result.

Have a photo, video or audio clip whose origin you want to check? Google opened SynthID Detector to everyone on 2026-10-07, with a globally available English interface, according to its official announcement. You can upload media at SynthID Detector and look for a supported invisible watermark.

The key limit: no watermark detected does not mean no AI was used. SynthID Detector checks for a specific provenance signal; it cannot certify every file as human-made or authentic.

What does SynthID Detector actually check?

SynthID embeds a signal directly into media made by compatible AI systems. People generally cannot see or hear it, but a matching detector can look for it. This differs from judging whether a face looks artificial or a recording sounds synthetic. Google DeepMind's explanation describes watermarking for images, video, audio and text; the public upload portal discussed here accepts images, video and audio. Do not confuse that with an open text-detection service.

Which companies are currently covered?

Google's announcement names Google, OpenAI, NVIDIA and Kakao. Apple is described as coming soon, rather than already supported. The provider list in the supplied screenshot matches the four currently named companies.

SynthID Detector screenshot listing NVIDIA, OpenAI, Google and Kakao — The author's original screenshot of the provider list. A listed company does not imply coverage of every product or historical output.

Treat this as a list of participating providers, not a guarantee for all their AI products. Coverage can depend on the product, model, creation date and export path. For example, OpenAI's documentation describes SynthID for supported image and audio outputs, while video uses Content Credentials. An OpenAI image result therefore does not establish that every OpenAI video can be detected here.

For concrete Google model examples, explore Nano Banana 2 for images and Veo 3.1 for video. These model pages explain the available generation workflows; a model link is not a guarantee that every APIXO export will be detected by SynthID.

Supported formats: more than photos

The author's upload screenshot lists these file formats:

Images: JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF, GIF
Video: MP4, MOV, WEBM
Audio: WAV, MP3, OGG, FLAC, AAC, M4A

SynthID Detector upload interface showing accepted image, video and audio formats — The author's original upload screenshot. File-format acceptance and watermark detection coverage are different.

A supported extension means the portal accepts that file type; it does not guarantee that every file of that type contains a detectable watermark. Check the live uploader for current size, duration and availability requirements.

How to test your own file

Open SynthID Detector in a new tab. Complete the site's sign-in steps if prompted; follow the current page instructions.
Choose a photo, video or audio file. Prefer the original exported file to a version repeatedly downloaded, cropped or compressed.
Drag the file into the upload area, or use the file-selection button. Wait for the analysis to finish.
Read the result and any provider label or warning together. Keep the file's original source and editing history in mind.

No sample yet? Generate and download one on APIXO

You can create an image, video or audio sample in the APIXO AI Generator, then download it and try it in SynthID Detector. For an audio example, explore Suno V6. The generation service and the detector serve different purposes: being able to generate a file does not guarantee a detectable SynthID watermark.

The supplied APIXO screenshot shows a Seedance 2.0 video result. It illustrates where to find a download, not a positive SynthID test or confirmed coverage for this model.

Open APIXO AI Generator, select image, video or audio and an available model, then create a sample.
When the result is ready, click Download on its result card to save the original file.
For an earlier result, click Recent, open a recent conversation and find the media you want. The supplied Recent screenshot says it shows conversations from the last 15 days; follow the current interface rather than assuming permanent access to all history.
Choose one or more media items from the conversation and download each relevant result. Return to SynthID Detector and test the saved files individually, keeping a record of the model and result.

APIXO video result with arrows pointing to the Download button and Recent conversations entry — The author's original APIXO screenshot shows the direct Download button and Recent entry. This is a download workflow, not a SynthID detection result.

APIXO Recent panel with an arrow pointing to a conversation and a notice about the last 15 days — The author's original Recent screenshot: open a conversation, then download the media you want from its results. The visible notice limits the displayed conversations to the last 15 days.

A real test: an image flagged as OpenAI media

To explore OpenAI image generation, see GPT-Image-2. The test screenshot below does not name its exact model, so it should not be presented as a test of that particular model.

SynthID Detector identifying an OpenAI watermark in an image of a red-haired painter — The author's real image test: the tool reports a detected watermark associated with OpenAI and warns that later edits may have occurred.

In the author's test, an image of a red-haired painter returned “SynthID was detected” and “This media was made or edited with OpenAI”. The screen also warns that further editing may have happened afterwards.

That is useful evidence of AI generation or editing associated with OpenAI. It does not tell us the exact model, prompt, creator or proportion generated by AI. This is one supplied image result, not an accuracy study or a video-and-audio benchmark.

Where the detector can miss AI content

Other providers or unsupported outputs: a file without a supported SynthID watermark can still be AI-generated. A familiar company name alone is insufficient to establish coverage.
Older files and export differences: media created before watermarking was enabled, or through an unsupported product, model or export path, may carry no usable signal.
Substantial changes: the watermark is designed to survive common edits, but extensive cropping, compression or other transformations can make it undetectable. Robustness is not a guarantee for every altered file.
A positive result has limits too: it identifies a provenance signal, not whether the depicted event happened or the caption is accurate. AI editing does not necessarily mean the whole file was generated from scratch.

OpenAI explains these provenance limitations. When no watermark is found, keep the question open: check the original source, compare other versions and look for available provenance information. Do not turn a negative result into a certificate of authenticity.

Have something you want to try?

Do you have a photo you would like to check? Or a video or audio clip? You can also try a file from another company's AI product—just treat it as an experiment with unknown coverage.

Record the originating product and model, if known, and whether you still have the original file.
Test the file and save the result, including a negative or inconclusive outcome.
If comparing an edited version, record the changes. One result does not establish a provider-wide detection rate.

Open SynthID Detector and test your file. Which images would you try first, and which other companies' AI outputs would you like to compare?

Official sources

Sources checked on 2026-10-09. Screenshots supplied by the author; only the image example has a supplied detection result.