Images from AI tools can carry records of how they were made. Some are plain text that anyone can read, some are signed provenance data, and some are invisible watermarks in the pixels. This article explains each kind, how Content Credentials (C2PA) work, and what they do and do not prove.
Four kinds of AI information
- Generation records. The prompt, negative prompt, seed, sampler and model. Stable Diffusion web interfaces save these as text in PNG files, and ComfyUI also saves the whole node graph, the workflow.
- Provenance records. Content Credentials, a signed manifest that says which tool made or edited the file.
- Source labels. Standard fields, such as the IPTC digital source type
trainedAlgorithmicMedia, that state the image was generated by a model. - Invisible watermarks. Signals woven into the pixels, such as Google’s SynthID. They are not metadata.
The first three live in the file’s metadata and can be inspected and removed. The fourth cannot be seen by a metadata tool at all.
How Content Credentials work
Content Credentials implement the open C2PA standard. When a tool creates or edits an image, it can attach a manifest to the file. The manifest contains:
- A claim naming the application that produced the file.
- Assertions, such as the actions taken (created, cropped, edited with a generative tool) and whether a model produced the content.
- A signature from a certificate that identifies the signer, so a viewer can tell who vouches for the claim.
- A binding to the image, a cryptographic hash of the pixel data. If the pixels change afterwards without a new signed record, the binding breaks and a verifier says so.
The manifest is stored in the file: in a caBX chunk in PNG, in APP11 segments in JPEG, in a C2PA chunk in WebP, and in a box in MP4 and MOV. A verifier such as the Content Credentials verify page reads it and shows the history.
What C2PA can and cannot prove
- It can show which tool claims to have created or edited a file, who signed that claim, and whether the pixels have changed since.
- It cannot show that what a picture depicts is true. A real photo can be staged and a generated image can be accurate.
- A missing manifest proves nothing. Plenty of editors, platforms and messaging apps drop metadata, and a manifest can simply be deleted, so an image without credentials is not necessarily real.
- A present manifest is a claim, not a verdict. It is only as trustworthy as the signer.
Some providers also keep a copy of the credentials online and can recover it from an invisible watermark or a fingerprint of the image. That is why platforms sometimes show a label even after the file’s own manifest is gone.
Who writes it and who reads it
Several image generators and editors attach Content Credentials by default, among them OpenAI’s image generation, Adobe Firefly and Photoshop’s generative tools, and Microsoft’s design tools. Some cameras and phones can sign photos as well. On the reading side, Meta, TikTok and LinkedIn have each said they read Content Credentials or IPTC fields to show labels, and YouTube asks creators to disclose realistic synthetic content when they upload.
Prompts and workflows in PNG files
Stable Diffusion interfaces such as AUTOMATIC1111 write one text entry called parameters into each PNG, holding the prompt, the negative prompt, steps, sampler, seed, size and model identifiers. ComfyUI writes two, prompt and workflow, and dragging such a PNG back into ComfyUI rebuilds the whole graph. That is handy when you share on purpose and a leak when you do not. The AI image metadata remover goes through what each tool writes.
Keep it or remove it?
There is no single answer, and the choice is yours.
- Keep it when provenance helps you: journalism, product and documentary photography, or any case where you want to show a file is what it claims to be. Also keep it when a platform or client requires it.
- Remove it when the records expose something private: prompts and workflows that are your work, client or product names, model and folder details, or a signer identity you do not want attached to a public file.
- Either way, keep the original so you retain the full record, and be honest about AI use where your audience or a platform expects it. Removing metadata does not change what an image is.
How to check an image
- Drop it on the AI metadata cleaner. The report names generation records, Content Credentials, XMP and comments if it finds them.
- Drop it on the Content Credentials verify page to read a manifest.
- Run
exiftool -a -G1 -s image.pngon a desktop to see text entries and labels.