How to Check an Image for an AI Watermark (SynthID) or C2PA

Images made with ChatGPT ship with their origin written into the file. Gemini's do too. The fastest way to check an image's provenance is to read every signal it carries: the embedded SynthID watermark, signed C2PA credentials, and IPTC disclosure metadata. Slop or Not reads all of it in one pass, and when the file carries nothing, it falls back to pixel-level detection. You can try the free AI watermark detector in any browser right now.

This guide covers what an AI watermark actually is, how it differs from C2PA and IPTC provenance, how to check each signal on an iPhone, a Mac, or the web, and what a blank result really tells you.

• What Is an AI Watermark?
• How Do I Check for an AI Watermark on iPhone or Mac?
• Can I Check for an AI Watermark Online?
• Which AI Generators Add Provenance Signals?
• What If There Is No Watermark?

What Is an AI Watermark?

AI images can carry several machine-readable provenance signals, but only some are watermarks. Google's SynthID is embedded in the pixels. C2PA Content Credentials and IPTC disclosure fields are metadata attached to the file, not pixel watermarks.

SynthID is the one built to last. Google DeepMind weaves it into the image data during generation, so you can't see it and a metadata cleaner can't reach it. There's nothing in the file's fields to clean. DeepMind says the signal holds up through the kind of handling images actually get, like resizing, mild compression, and format conversion. That's why a positive SynthID read on a much-shared image still means something.

Metadata can say more, but it is fragile. A C2PA Content Credential is a signed manifest. Reading it shows the assertion stored in the file, while cryptographic validation checks that the signed data is intact. A trust check asks whether the signer is recognized. Only then can you interpret whether the assertion records generation, editing, or some other processing; C2PA alone does not prove that the pixels are AI-generated. IPTC digital-source fields carry a plainer, unsigned disclosure that an editor can change later.

And metadata disappears constantly. Take a screenshot and you've made a brand-new file with no history. Post an image to most social apps and the re-encode quietly drops whatever records were attached. In practice, files with no C2PA manifest are the norm, so missing metadata proves almost nothing on its own. We wrote about what an AI metadata remover actually hides if you want the full story on stripped files.

How Do I Check for an AI Watermark on iPhone or Mac?

Three taps, more or less:

1. Tap Share on the image, wherever you found it. Photos, Safari, Messages, Files, they all have the button.
2. Pick Slop or Not in the share sheet.
3. Read the verdict. If the file carries SynthID, a C2PA credential, or IPTC disclosure fields, the result shows the evidence it found, including the credential and claimed source.

That's the entire check. It runs on-device, your photo is not uploaded anywhere, and there's no account to create. On a Mac with Apple silicon the same check runs natively; detection uses Apple silicon accelerated by the Neural Engine on iPhone and the Neural Engine or GPU on Mac, so a verdict lands in seconds, even with no connection at all.

The result screen shows the probability alongside whatever provenance evidence the app read, so when you have to explain a call to an editor, or to a student, you can point at the specific signal it found instead of a hunch. A journalist can verify a source photo without it ever leaving the phone, and a teacher can check what a student handed in without shipping student work to a third party. Download Slop or Not free for iPhone and Mac.

Can I Check for an AI Watermark Online?

Yes. The free online AI image detector runs the same layered check from any browser, which makes it the route for Windows, Android, Linux, and ChromeOS. One thing to be clear about: online checks are not on-device. The image goes to Numen's private Mac server for processing and is deleted after processing. No account needed, nothing to install, and it's a quick way to run a handful of images back to back from a desktop browser.

If you already have a suspect in mind, the focused tools give you a sharper starting point. The SynthID watermark checker reads the embedded signal Gemini images carry. The C2PA Content Credentials checker reads signed manifests, validates signatures from recognized issuers, and reports the source it recognizes. Same server, same delete-after-processing rule.

Pick the surface based on how sensitive the image is. A meme from the group chat can go to the browser. A file you wouldn't email to a stranger belongs in the native app, where it never leaves your hardware.

Which AI Generators Add Provenance Signals?

Fewer than you'd hope. Six generators add provenance signals to their output today, using pixel watermarks, signed credentials, or disclosure metadata. OpenAI's image generation, including ChatGPT, attaches C2PA Content Credentials and embeds SynthID, and OpenAI documents both. Google Gemini, including Nano Banana, embeds SynthID. Adobe Firefly attaches C2PA Content Credentials, which the Adobe Firefly image checker reads directly. Microsoft Designer writes C2PA and IPTC metadata. Apple's Image Playground writes an IPTC credit that names it. Anthropic says supported files Claude produces carry C2PA Content Credentials; its text watermark is a separate SynthID-Text mechanism planned for future Claude models.

But signed evidence runs out fast. Midjourney typically ships no watermark Slop or Not can verify, and images from Stable Diffusion or Flux carry no embedded watermark and no signed manifest either. Open-source pipelines can add a watermark, but the step is optional and most workflows skip it.

The apps people run those models in leave more behind. Stable Diffusion WebUI, ComfyUI, InvokeAI, Draw Things, and NovelAI all tend to write their generation settings into the file, and Slop or Not reads that layer too, naming the tool that made the image. It is a weaker class of evidence than a signed credential. Nobody signed a parameter block and any editor can delete it, so a match tells you something and a blank tells you nothing.

A Firefly credential becomes strong provenance evidence after the manifest is read, its signature validates, the signer is trusted, and the generation or editing assertion is interpreted. When those checks line up, it is evidence of Firefly involvement, not proof that every pixel is AI-generated or unchanged. A blank result on something that might have come from Midjourney is not an answer, because the check was looking for evidence that was never added. For the two generators with the strongest signals, our walkthrough of how Slop or Not verifies OpenAI and Gemini images with SynthID shows what a positive read looks like in practice.

What If There Is No Watermark?

A missing watermark is not proof that a human made the image. Screenshots and social-app re-encodes strip metadata from real photos and AI images alike, and most open-model generators never added anything verifiable to begin with. All a blank result tells you is that the provenance layers had nothing to say.

So the pixel layer takes over. When the watermark and metadata checks come up empty, Slop or Not's on-device classifier examines the image itself and reports a probability. In our internal tests it flags AI-generated images at about 90% accuracy. A supported SynthID read or a trusted, validated C2PA assertion adds provenance evidence; neither turns the pixel score into certainty.

Accuracy is based on our internal tests. Results can vary with new AI models and advanced methods designed to trick detectors.

One practical habit helps more than anything else here: get the original file when you can. A screenshot of a suspect image throws away whatever provenance the original was carrying, and the classifier ends up working from a re-encoded copy. The original gives every layer more to read.

Read the verdict the way the app presents it, as evidence. A high probability on a stripped file is a good reason to ask where the image came from. With C2PA, a readable manifest tells you what the file claims, cryptographic validation shows whether the signed record is intact, and the trust check identifies whether the signer is recognized. You still have to interpret the assertion. A trusted assertion that an AI tool generated or edited the file is strong provenance evidence, but it does not prove that every pixel was AI-generated or that no later edits occurred.