Which Apps Write AI Metadata Into Your Photos
Some write a signed credential, some write your entire prompt, and one of the most popular writes nothing at all. A reference for what ends up in the file.
Three different kinds of record
AI metadata is discussed as one thing and is really three, differing in what they contain and how much they give away.
- Signed provenance manifests. C2PA Content Credentials — a cryptographically signed statement of which tool produced or edited the image. Designed to be read by platforms.
- Generation parameters. Plain text blocks holding the prompt, negative prompt, seed, sampler, step count and model hash. Designed for the user, and readable by anyone with the file.
- Generator tags. A tool name dropped into an ordinary field such as Software or CreatorTool. Incidental rather than deliberate.
Adobe Firefly, Photoshop and Lightroom
Adobe is a founding member of the C2PA coalition, and Firefly has embedded Content Credentials since launch. Every Firefly output carries a manifest identifying it as generated.
Photoshop and Lightroom write credentials when generative features are involved. Since 2026 this is mandatory for such workflows, with the in-app option to disable it removed, and photographers have reported credentials appearing on exports more broadly than they expected.
This is the source of most surprise labels on real photographs, because it is the only entry on this list that routinely attaches to images that are overwhelmingly camera-captured.
OpenAI: DALL-E 3, ChatGPT images and Sora
OpenAI added C2PA Content Credentials to DALL-E 3 output in 2023, and extends the same approach to images generated in ChatGPT and to Sora video.
The manifest states that the image was generated by an OpenAI model. It does not carry your prompt, which is a meaningful distinction from the Stable Diffusion approach below — a DALL-E file announces what made it without disclosing what you asked for.
Google: Gemini and Imagen
Google output carries two separate things, and conflating them causes real confusion.
The first is Content Credentials, the same C2PA manifest the other tools write. Gemini image models have carried these alongside their other marking since late 2025, and it is ordinary metadata that can be removed like any other.
The second is SynthID, an invisible watermark embedded in the pixels themselves. It is not metadata, no metadata tool removes it, and it is built to survive cropping, resizing and re-compression. Any tool claiming to strip SynthID is describing something it cannot do.
Stable Diffusion and ComfyUI: the ones that reveal most
These write no C2PA manifest, which sounds like less exposure and is in practice far more.
Stable Diffusion interfaces write a PNG text chunk holding the full generation record: prompt, negative prompt, seed, sampler, CFG scale, step count and model hash. Anyone with the file can read all of it in a text editor.
ComfyUI goes further and embeds the entire node graph as JSON — your complete workflow, node by node, including every model and setting. It can add tens of kilobytes to each image.
For anyone whose prompting or workflow is part of how they work, this is the most consequential metadata on this page. A platform label tells viewers a tool was used. A parameters chunk hands them the method.
Midjourney: the gap in the system
As of early 2026 Midjourney does not embed C2PA Content Credentials. Files may still carry ordinary metadata from wherever they were saved or converted, but the signed manifest platforms look for is absent.
This produces the asymmetry described elsewhere on this site: a photographer who removed a power line in Photoshop gets a platform label, while a fully generated Midjourney image does not. The system marks the tools that chose to participate.
It is a reason to read a missing label as weak evidence. Absence of a manifest says nothing about how an image was made.
How to find out rather than assume
This list will age. Tools add credentials, change defaults and rename fields, and a page like this is a snapshot rather than a standing truth.
The durable approach is to inspect an actual file from your own workflow. Export as you normally would, then look at what is in it: whether a C2PA manifest is present, whether a parameters or workflow chunk was written, and which fields name a tool.
That takes a minute and answers the question for the version you are actually running, which is the only answer that matters.
Inspect a file from your workflow
See exactly which AI markers an image carries before you publish it. This tool only reads — nothing is changed, and nothing is uploaded.
Open AI Metadata CheckerFrequently asked questions
Which AI image tools embed C2PA Content Credentials?
Adobe Firefly since launch, OpenAI DALL-E 3 and ChatGPT images since 2023, and Google Gemini and Imagen models. Adobe Photoshop and Lightroom also write them when generative features are used. As of early 2026 Midjourney does not.
Which tools reveal my actual prompt?
Stable Diffusion interfaces write the prompt, negative prompt, seed, sampler and model hash into a PNG text chunk, and ComfyUI embeds the entire node graph as JSON. Both are plain text that anyone with the file can read. C2PA manifests from Firefly and DALL-E do not include your prompt.
Does Photoshop add AI metadata to photos that are not AI-generated?
It writes Content Credentials when generative features are involved, which includes small retouching edits on an otherwise camera-captured photograph. Since 2026 the in-app opt-out for such workflows has been removed, and photographers report credentials on exports more broadly than expected.
What is the difference between SynthID and Content Credentials?
Content Credentials are metadata — a signed manifest stored alongside the image data, removable like any other metadata. SynthID is a watermark embedded in the pixels themselves, designed to survive cropping and re-compression, and no metadata tool removes it.
Why does a ComfyUI image have such a large file size?
Because the full workflow is embedded as JSON. A complex node graph can add tens of kilobytes to every exported image, on top of the picture itself. Removing that block shrinks the file with no effect on image quality.
If an image has no AI metadata, was it made by a camera?
No. Some generators write nothing a platform reads, and metadata can be stripped by any number of ordinary steps including screenshots, conversions and platform re-encoding. Absence of a marker is not evidence of origin.
Do these tools write metadata I cannot remove?
Only the pixel watermarks. C2PA manifests, parameter chunks, workflow JSON and generator tags are all metadata and all removable losslessly. SynthID and similar in-pixel marks are not, and this site will not claim otherwise.
How can I check what my own workflow writes?
Export a file exactly as you normally would, then inspect it. Tool behaviour changes between versions and defaults shift, so testing your own export is more reliable than any list, including this one.