Check if a photo, clip, or audio is AI with SynthID Detector

Upload the best-quality copy you have to SynthID.com. Detected means a SynthID-enabled model probably made or edited the file. Not detected does not mean a person made it.

SynthID Detector is the public page for that check. It looks for one invisible watermark. It does not grade how AI-like a picture feels. The lines below are from SynthID.com and its FAQ, checked October 16, 2026, plus the Google launch post and the Ars Technica report already linked from the tool page.

See which labs leave a mark

The FAQ opens with the definition.

"SynthID is a digital watermarking approach created by Google DeepMind."

The same answer says the mark is woven into the pixels of an image or a video, and into the frequency components of audio. It also covers text. A logo in the corner can be cropped off. File metadata can be stripped. This mark is meant to travel with the content. The FAQ then says no signal is completely robust, and it tells you to collect more than one point of data before you decide. Hold that sentence. The rest of the page is an application of it.

Who embeds the mark. The FAQ names the labs in one line.

"SynthID is an approach for watermarking and identifying AI-generated content generated by models across Google, Nvidia, OpenAI, and Kakao."

On the detector those four are labeled Google AI, NVIDIA, OpenAI, and Kakao. If a result names a partner the page does not have a logo for, the label it falls back to is "AI from a SynthID partner." The FAQ says the list grows as more organizations adopt the watermark. Today the named set is those four.

The detector page states the limit in its own sentence, next to the upload box.

"This is not a general AI detector. We can only detect media from the above companies, who have adopted SynthID technology."

The FAQ draws the same line under "What is the difference between SynthID and AI detection tools?" SynthID looks for a signal that was embedded on purpose when the file was made. A general detector guesses from leftover patterns when no signal was embedded. The FAQ's own limit on the watermark path:

"It provides high reliability and resilience, though it is limited to identifying content generated by SynthID-integrated models."

A picture from a lab that never joined, or from a model someone ran at home, is outside this check. The page is not offering to become a different tool.

Google opened the public detector on October 7, 2026. Pushmeet Kohli's post that day, We're making it easier to identify AI-generated content globally, says access is expanding to everyone, "with the tool available globally in English starting today."

Upload the best file you have

Open SynthID.com. The page says "Upload a file for detection." Drag the file in, or select it. You do not choose a media type by hand. The page reads the file and routes an image, a video, or an audio file to the matching check.

If you are not signed in, the upload stops. The dialog says "Please sign in before detection." The buttons are "Sign in with Google," "Sign in with Apple," and "Sign in with ChatGPT." Ars Technica, on October 7, 2026, describes that gate as Google, OpenAI, and Apple accounts. The ChatGPT button is the OpenAI login.

Use is capped. Ars quotes Google directly: each signed-in user gets "a daily usage quota of approximately 10 image, video, and audio checks." That quota is approximately 10 a day.

The FAQ lists three kinds of file. The result is not the same shape for each.

• Images. The check looks for the presence of a watermark in the image.
• Video. The check looks for segments of the video with the watermark.
• Audio. The check looks for segments of the audio with the watermark.

A photo comes back as a read on the whole file. A clip or a recording points at the watermarked stretches. If only part of the video was touched, that part is what the FAQ says the result is about.

The Media button is labeled "Media: Supported file types." A file the page does not recognize gets the notice "Unsupported file type! Supported formats:" and then the same lists. These are the extensions, from the page checked October 16, 2026. The page prints them in capitals.

Kind	Extensions the check accepts
Image	jpg, jpeg, png, bmp, webp, avif, heic, heif, tiff, tif, gif
Audio	wav, mp3, ogg, flac, aac, m4a
Video	mp4, mov, webm

An empty file gets "File is empty." A file over 104,857,600 bytes gets "File is too large." That cap is 100 MB, and the number is the same for an image, a video, and an audio file.

Quality beats the extension. The detector page says so.

"When uploading files for detection, please upload the highest quality file possible to ensure maximum detection accuracy."

The FAQ says it again in the results section: "Use high-quality media files for best results, as heavy compression and editing impact detection."

Send the original export. A screenshot of a screenshot is a worse file for this check. So is a hard crop, or the copy a chat app recompressed on the way through. Those are edits and compression, which is the failure the FAQ names. If you still have the download, the camera-roll original, or the file the app saved, that is the one to upload.

Text is the odd one out. The "What is SynthID?" answer includes watermarks in text. The detector's supported types are the three in the table. A document is not one of them. Dropping one on the page is not a result.

Read detected and not detected

The FAQ section "How do I interpret detection results?" is the caption. Use those words when you tell someone else what came back.

"Watermark detected: A SynthID watermark is present in the content, and it was likely edited or generated by a SynthID-enabled AI model."

Detected means a SynthID-enabled model probably made the file, or edited it. The sentence covers both. The page does not split them. A file that is entirely generated and a real photo that was only lightly touched up with a SynthID-enabled tool can both come back detected. This result cannot tell those two apart. Ars Technica reports the same limit from the launch: the site tells you whether SynthID was used to produce or edit the media, and it stops there.

"Watermark not detected: No confident signal was found, so it was not likely to have been edited or generated by a SynthID-enabled AI model. Although this does not rule out AI generation or manipulation."

The clause after the period is the one to keep. Not detected does not mean a person made it. It means this watermark was not found with confidence. The file can still be AI-made. It can still be a real photo that was edited. The FAQ says so.

FAQ label	What you can say	What you cannot say
Watermark detected	A SynthID-enabled model likely made or edited this file.	That it is fully AI-made, or that a light touch-up is ruled out.
Watermark not detected	No confident SynthID signal was found.	That a person made it, or that no AI was involved.

Expect a miss, and a rare false hit

The FAQ names two reasons an AI-made file can still come back not detected.

"It was created by a model without SynthID (non-partner or pre-adoption)."

Non-partner is a lab that never joined. Pre-adoption is a file made before a lab that did join had turned the watermark on. A partner's name on the page does not reach backward through that lab's older output. If the photo, clip, or audio predates the adoption, a miss is the result the FAQ describes.

"Heavy modifications degraded the watermark signal."

The robustness answer names filters, color changes, and lossy saves. It says the mark is designed to stay detectable after those. It also says the watermark is not infallible. If someone tries many transformations, they may find one that does not get detected. Heavy compression and a stack of edits are the practical version of that sentence. Upload the best copy because of this, not because a bigger file looks more serious.

False hits are in the FAQ, and the bound is not a percentage.

"In very rare cases SynthID can also falsely trigger on non-watermarked content."

A detected result is a real signal. Very rare is the phrase the FAQ uses. The FAQ prints no accuracy rate, and none is added here.

Gather more evidence after a no

The FAQ does not tell you to stop at the badge. Two lines say what comes next.

"However, no signal is completely robust and we recommend collecting multiple points of data before making a decision."

The results answer says "gathering additional evidence is recommended."

A no is a hint. Take it to the next thing you already have: the account that posted the file, a second copy that was not recompressed, the context around the post. The detector has finished when it tells you whether this watermark is present. Proving a person held the camera is a different task, and the FAQ does not take it on.

Open SynthID.com, sign in, and upload the best copy you have. Read the two FAQ labels before you repeat the result. The directory note is /tools/synthid-detector.