Glossary

An image is pixels plus metadata. The pixels decide what you see; the metadata records who, when, with what and how. Deleting metadata does not change the picture.

Coalition for Content Provenance and Authenticity, a standard pushed by Adobe, Microsoft, the BBC and others. It records who created the content, with which tool and what edits were made, signs that record, and embeds it in the file. A platform that reads it has grounds to judge whether the content was AI-generated.

Adobe's commercial name for a C2PA manifest. Seeing “Content Credentials” is essentially the same as seeing “C2PA manifest”.

JPEG Universal Metadata Box Format. A C2PA manifest is stored as a JUMBF structure: an APP11 segment in JPEG, a caBX chunk in PNG, a uuid box in MP4. Cleaning C2PA means deleting these containers.

Extensible Metadata Platform, led by Adobe. Generators favour it for prompts, seeds and model IDs because the structure is flexible and costs almost nothing to include. Platforms also find AI tool names in it easily.

Exchangeable Image File Format. Records body, lens, exposure, time, GPS, and the software that wrote it. EXIF straight out of a camera is normal; the trouble is that the Software field can name an AI editor outright.

A standard from the International Press Telecommunications Council. Holds title, author, keywords and copyright. Adobe tools write their AI attribution fields here.

PNG allows arbitrary text chunks alongside the image data. Stable Diffusion and ComfyUI stuff the full prompt, the sampling parameters and often the entire workflow JSON into them — tens to hundreds of KB is common.

MP4 / MOV store title, author, comment and encoding software in the udta and meta boxes under moov. Removing them does not require re-encoding the video.

An invisible watermark embedded directly into the pixel data, verifiable only with Google's private key. Because it is fused with the pixels, removing it means repainting the image — a metadata tool cannot touch it at all.

How Instagram and Facebook mark suspected AI content. The trigger is mainly the C2PA and XMP in the file, not a per-pixel judgement. That mechanism is exactly why real photographs get mislabelled.

Tools such as Photoshop use AI to repaint part of an image. For the photo as a whole it changed one small area, but the export writes C2PA — and the platform does not distinguish a local retouch from a fully generated image.

When a genuinely taken photograph carries provenance data because it went through an AI-assisted edit (Generative Fill, AI denoise, AI background replacement) and is therefore labelled AI. This is the main real use of tools like this one.

An identifier encoded into tiny variations of the pixels, invisible to the eye. The fundamental difference from metadata is that it is part of the image content, so removing it necessarily changes the image.

Rewriting only the metadata blocks in the file header while copying the image data byte for byte. No encoder is involved, so quality loss is not structurally possible. This is the path taken for JPG / PNG / WebP.

Decode the image to pixels, then encode it again. The pixels go through one extra compression and the format may change. It is done this way because those two containers bind metadata to structures the decoder needs, so it cannot be removed in place.