How to check AI image metadata

If you want to know what an image says about its own origin, there are a handful of specific places to look. Here is what they are and how to read them.

The five places to look

1. IPTC DigitalSourceType. The standardised vocabulary term for how media was produced. The value trainedAlgorithmicMedia means “generated by an AI model”, and compositeWithTrainedAlgorithmicMedia means AI elements were combined with other content. This is the closest thing to an industry-standard “this is AI” flag, and it lives in the XMP packet.

2. C2PA manifests. A signed provenance record naming the generating tool, stored in an APP11 segment (JPEG), a caBX chunk (PNG) or a dedicated chunk (WebP). See What is C2PA?.

3. Software and CreatorTool fields. Ordinary EXIF and XMP fields naming the application. Often the plainest signal available — a value like “ChatGPT” or “Adobe Firefly” needs no interpretation.

4. PNG text chunks. The richest source by far, and the one people are least aware of. Stable Diffusion interfaces write the full prompt, negative prompt, seed, sampler, step count and model hash into a chunk called parameters. ComfyUI stores an entire workflow graph under prompt and workflow. InvokeAI and NovelAI have their own conventions. These chunks can be compressed, so a reader has to inflate them to see the contents.

5. Generator-specific XMP namespaces. Some tools declare custom namespaces that identify them even when the obvious fields are empty.

Doing it

Open the AI Watermark Checker and drop your image in. It reads all five and groups the results: AI provenance, private information, metadata blocks, and hidden characters found inside the metadata itself.

Expand any row to see what that signal is, what it means, and whether it can be removed. The values are shown as they actually appear in the file — sanitised for display, since metadata is untrusted input and can contain control characters designed to mislead.

Everything happens in your browser. Disconnect from the internet after the page loads if you want to confirm it.

Reading the results honestly

Three statuses matter, and the difference between them is the point.

Detected means the signal was found in the file.

Not detected means we looked in the places that signal lives and it was not there. For metadata, this is a meaningful negative — we can genuinely establish that a manifest is absent from a file.

Unable to verify means we have no way to look at all. SynthID reads this on every image, permanently, because detecting it requires Google’s own detector. This is not a weaker form of “not detected” — it is a statement about our instruments rather than about your file.

A tool that collapses those last two into one green badge is telling you something it does not know.

The interpretation trap

The most common mistake is treating missing metadata as evidence of human authorship. It is not, even slightly.

Metadata is lost through completely ordinary activity: social platforms re-encode uploads and strip it, messaging apps compress and strip it, screenshots discard it entirely, and editors without C2PA support drop manifests on save.

So the population of AI-generated images carrying no AI metadata is vast, and mostly nobody removed anything deliberately. Finding provenance data is informative. Not finding it is close to uninformative.

If you want it gone

The AI Metadata Remover strips all of it — prompts included — and re-scans its own output so you can confirm the prompt text is genuinely no longer anywhere in the file.