What each AI generator embeds in your files
Not every generator marks its output the same way. Some attach a signed C2PA manifest, some write plain text chunks, and some do very little. If you are handing files to clients, it is worth knowing which category yours falls into.
Generators change their behaviour between versions, and this can shift without announcement. Treat the notes below as orientation, then scan one of your own exports to see what your current setup actually produces.
Midjourney
Midjourney output is typically PNG, and PNG text chunks are the thing to watch. These chunks are plain, uncompressed key-value text and have historically been used to carry prompt text, job identifiers and generator attribution.
This is the case where scanning your own files is most worthwhile, because prompt text embedded in a deliverable is the kind of thing you only discover after a client mentions it.
DALL-E 3 and ChatGPT images
OpenAI attaches C2PA Content Credentials to DALL-E 3 output. That means a signed manifest declaring the content as AI-generated, stored in a JUMBF box in the file. Depending on the delivery format you may also see EXIF software attribution alongside it.
Because the manifest is signed, it is not something you can edit convincingly — it is either present and valid, or absent.
Runway
Runway exports carry a visible corner watermark on many plan tiers, plus container metadata identifying the generator. The visible mark and the embedded metadata are separate problems: removing one does nothing about the other.
Video is also where in-browser tooling hits its ceiling, since removing a visible watermark from a clip requires re-encoding every frame.
OpenAI Sora
Sora follows the same C2PA approach as DALL-E, with a manifest declaring AI generation, alongside a visible watermark on video output. Metadata inspection works on the container regardless of what the visible watermark does.
Stable Diffusion, ComfyUI and Automatic1111
Local generation pipelines are the most revealing of all, because they are built for reproducibility. PNG text chunks commonly carry the full generation parameters: prompt, negative prompt, seed, sampler, CFG scale, model hash and more.
That is genuinely useful while you are iterating, and genuinely awkward if it ships to a client. It is worth stripping deliberately rather than by accident.
Adobe Firefly
Adobe co-founded the C2PA effort, so Firefly output carries Content Credentials by design, and Adobe applications are among the few that display and verify them natively.
The practical advice
Rather than memorising this, build one habit: before a file goes to a client, scan it. It takes a second, it is free, and it tells you what is actually in that specific file rather than what a generator did six months ago.