Photo Verification

SynthID Checker: How to Check an Image for the Watermark, and What the Result Proves

Google's SynthID Detector is open to everyone at synthid.com since October 2026. How to check an image, video or audio file for the watermark, how to read a hit or a miss, and what none of the checks can tell you about a photograph.

ByLumethic Team
••Updated •7 min read
Share

To check an image for SynthID, go to synthid.com, sign in, and upload the file. Google opened its SynthID Detector to everyone on 7 October 2026, so the waitlist that used to stand in front of it is gone. The same check also runs in the Gemini app and in "About this image" in Google Search and Chrome, and OpenAI's Verify tool reads both marks OpenAI applies to a ChatGPT image. For the other half of the answer, the signed Content Credentials that name the generator outright, our free AI Photo Checker reads them in your browser with no upload. Every one of these checks answers one question well, whether a participating AI generator made the file, and leaves the harder question open. This article covers how the detector works, how to read what it tells you, and where its answer stops. For the broader picture of who applies SynthID in the first place, start with our SynthID adoption overview.

What the Detector Is

Google announced the SynthID Detector at I/O as a standalone verification surface and ran it through a waitlist for journalists and researchers first. On 7 October 2026 it opened the tool to everyone, worldwide and in English, at synthid.com. The detector accepts direct uploads of images, video and audio and reports whether it finds the SynthID watermark that Google's generators and its partners embed at generation time.

Two things distinguish it from the older routes. It handles more than images, which matters now that SynthID marks audio and video from several vendors. And it is a dedicated page with an upload box, rather than a question you ask a chatbot or a menu you find behind a right-click.

The scale behind it is what makes the check meaningful. Google says it has watermarked more than 180 billion images and videos and 240,000 years of audio, and since May 2026 the mark also lands in every image from ChatGPT and the OpenAI API. Google names OpenAI, NVIDIA, Kakao and Apple as partners whose content the detector identifies, and ElevenLabs has committed to adding the mark to its audio.

How to Use It

Open synthid.com, sign in with a Google, OpenAI or Apple account, and upload the file. The check is free. The result says whether the watermark was found in the file, and for video and audio whether it was found in part of it.

There is no software to install and no API for the public tool. The classifier runs on Google's servers, which means the file leaves your machine. For a confidential image that is a consideration in itself, and one reason to read the Content Credentials first with a tool that stays in the browser.

How to Read a Result

A positive result means the detector found the statistical pattern SynthID embeds at generation time. For an image, that is strong evidence the file came from a participating generator or passed through a participating editor. Google's own limits page is explicit on the second case: the tool does not tell you whether a file was generated outright or only edited with an AI tool, and it does not say which parts are synthetic. A real photograph that went through Google's sky cleanup carries the same mark as a scene invented from a prompt.

A negative result is where reading discipline matters. The detector reports that it found no SynthID watermark. That is all it reports. The file may come from a generator that never participated in watermarking, the mark may have been degraded by heavy cropping, re-encoding or screenshotting, or the content may be a genuine photograph. The detector cannot distinguish between those three cases, and treating "no watermark" as "not AI" is the single most common misreading of any detection tool.

Detection is also probabilistic at the edges. SynthID survives normal compression and resizing well, and Google engineered it to be robust, but robustness has limits and the classifier returns a likelihood, not a certainty.

What It Cannot Tell You

The detector checks for one mark: SynthID. Its coverage is exactly the SynthID ecosystem, which as of October 2026 means Google's generators, OpenAI's images, and content from NVIDIA, Kakao, Apple and ElevenLabs. Images from Midjourney, Stable Diffusion, Flux and every locally run model carry no SynthID and never will unless those vendors join. Our generator-by-generator comparison tracks which tools mark their output and how.

It also cannot authenticate a photograph. Finding no watermark in a real photo is the expected outcome, but it is the same outcome as an unmarked fake. The detector's design goal is catching participating generators' output, not proving a capture happened. That distinction, detection versus provenance, is the subject of our comparison of the two approaches, and it is why a serious authenticity claim needs evidence that starts at the camera rather than a scan that ends at one vendor's watermark.

The Other Ways to Check

The same detection is reachable three more ways. The Gemini app runs SynthID checks on uploaded images when you ask whether an image is AI. "About this image" in Google Search and Chrome exposes the check for images Google can fetch, without an upload. OpenAI's Verify tool reads the two marks OpenAI applies to its images, SynthID and the C2PA manifest, and since July 2026 accepts audio as well.

The C2PA half you can read yourself. Our free AI Photo Checker parses Content Credentials directly in your browser, with no account and no upload to anyone's server. When a ChatGPT image still has its metadata, the manifest names the generator outright, which is a plainer statement than a watermark likelihood. When metadata is stripped, which platforms do routinely, the pixel watermark is what remains, and that check belongs to Google's and OpenAI's surfaces.

When the Detector Is the Wrong Tool

Use the detector when the question is "did this come from a participating AI generator." Use something else when the question is "is this photograph real."

The second question comes up when a photo is disputed: a contest entry, a news submission, an insurance claim, a marketplace listing. There, absence of a watermark proves nothing, and AI detectors that guess from pixels alone flag real photos often enough to make accusations unreliable. What settles the question is provenance: whether an original capture exists and whether the disputed image is consistent with it. That comparison, original against published copy, is what Lumethic's verification runs, and it works for photos regardless of which generator ecosystems exist around them.

The two approaches are complements, not rivals. A newsroom triaging a suspicious inbound image reasonably runs the watermark check first, because a positive hit ends the discussion in seconds. When the check comes back empty, the real work starts, and that work is provenance.

Frequently Asked Questions

Is there a SynthID checker I can use right now?

Yes. Google's SynthID Detector at synthid.com is open to everyone since 7 October 2026: sign in with a Google, OpenAI or Apple account and upload an image, video or audio file. The Gemini app and "About this image" in Google Search and Chrome run the same check, and OpenAI's Verify tool checks images from ChatGPT and the OpenAI API. None of these is a download; the classifier runs on Google's or OpenAI's servers.

How do I check an image for SynthID?

Upload it at synthid.com, or ask the Gemini app whether it was made with AI, or right-click it in Chrome and open "About this image". Each reports a found watermark. Then read the Content Credentials with our AI Photo Checker: when the metadata is intact, the manifest names the generator, which is a plainer statement than a watermark likelihood.

Does OpenAI use SynthID?

Yes. Every image from ChatGPT, Codex and the OpenAI API carries SynthID since 19 May 2026, together with a C2PA manifest, and OpenAI kept both marks on the ChatGPT Images 2.6 models released on 4 November 2026. Images made before May 2026, and images from generators such as Midjourney or Stable Diffusion, carry no SynthID mark.

Is the SynthID Detector available to everyone?

Yes, since 7 October 2026. Google opened it worldwide in English after a waitlist period that reviewed journalists, media professionals and researchers first. You need to sign in with a Google, OpenAI or Apple account.

Does the SynthID Detector find ChatGPT images?

Yes. OpenAI applies SynthID to every image from ChatGPT and its API since 19 May 2026, and the detector finds the same mark regardless of which partner embedded it.

Does it detect Midjourney or Stable Diffusion images?

No. Those generators do not apply SynthID. A Midjourney image comes back with no watermark found, the same result as a real photograph.

If the detector finds no watermark, is the image real?

No. The result means only that no SynthID mark was found. The image could be from a non-participating generator, the mark could have been destroyed by processing, or the image could be genuine. Establishing that a photo is real requires provenance evidence, not the absence of one vendor's watermark.

Does a SynthID hit mean the whole image is AI-generated?

No. Google says the detector cannot tell a generated file from one that was only edited with an AI tool. A real photograph cleaned up in Google's editor carries the same mark as a fully synthetic image. What the hit establishes is that a participating tool touched the file, and the rest of the question is answered by the original capture, not by the watermark.

Can SynthID be removed from an image?

Google designed it to survive normal handling like compression, moderate cropping and filters. Determined adversarial processing can degrade it, which is one more reason a missing watermark is weak evidence of anything.


Related Reading

#SynthID#AI Detection#Google#Watermarking#Provenance