AI Image Metadata, Content Credentials, and Watermarks Explained
Learn where AI-related information can live in an image and what a local metadata check can tell you.
An image exported from a generator or editor can carry clues about its history. A prompt might be stored in a text field. A signed record might describe an edit. A watermark might be embedded in the image itself. Each signal has a different location and answers a different question. Knowing which one you are looking at helps you protect private information and interpret the result of an inspection.
Suppose you are sending a concept illustration to a client. The artwork shows a generic product, but the prompt used an internal product name. You may want to remove that private text while preserving the original in your project archive. If the recipient needs a record of the image's origin, decide how to provide that before making a sharing copy. Inspect the exported file so your decision is based on what it actually contains.
What AI Image Metadata Can Reveal
AI image metadata can include a tool name, a prompt, generation settings, or workflow information stored alongside the image. What appears depends on the application, file format, and export settings. One tool might save a structured workflow in a PNG text record. Another might include a brief software label or a description in a field that a conventional photo editor can also write.
Think about what a prompt may contain beyond the obvious visual direction. You might include a client name, a description of an unreleased campaign, or a source reference. A workflow can also include model names and configuration values that tell someone how a result was made. Those details can be useful for reproducing your own work. An external recipient may only need the finished image and a short description of its use.
A software field deserves context. A label naming an editor can indicate that the application touched the file. It may say little about the parts of the image the application created, changed, or left intact. For example, a designer can open a camera photo in an application that offers AI tools and use it only to crop the image. Read the value and the field type before drawing a conclusion.
Ordinary photo metadata can matter too. An AI-assisted edit may start with an original photo that carries a location, a photographer's name, or a device model. If that information survives the export, it can still be relevant to the sharing decision. Inspect the rest of the file even when your first question is about a prompt or a generation tool.
What Content Credentials Add
Content Credentials use the C2PA standard to associate signed provenance information with an asset. A credential can record information such as a source, an editing action, or an ingredient used in the workflow. The available record depends on what participating tools recorded and what the asset still carries. See the C2PA FAQ for the distinction between provenance, validation, and claims about an asset.
Start by separating discovery from verification. Finding an embedded credential tells you a record is present. A credential verifier can check the signature and report how the record relates to the file. Read the verifier's result and its information about the signer. A signed claim can establish what a signer asserted without establishing the truth of every scene, caption, or statement associated with the image.
A photo captured by a camera can carry Content Credentials. So can an image edited in software or created with a generator. If a credential is present, read its recorded actions before interpreting what it says about AI use. When an image passes through several tools, the available provenance can represent only part of the journey. A client or editor who needs a fuller history may also need the original asset and a description of the workflow.
If you need the signed history, preserve the original file. Editing or stripping an embedded record can change what a verifier sees. Deliver the original through the route agreed with your recipient, or provide a separate distribution copy when that is appropriate. Keep clear filenames so that a privacy-cleaned export and a provenance-bearing original are easy to distinguish.
Where Visible and Invisible Watermarks Fit
A visible watermark is part of what you see when you open the image, such as a logo, credit line, or disclosure overlaid on the artwork. If it is baked into the exported pixels, it remains when the file's text fields are cleaned. Look at the image itself for notices, logos, or other content that should stay with the work.
An imperceptible watermark can be embedded as a signal in the media. A compatible detector checks for that particular signal. Removing an embedded text field or record is a different operation. For an example of the two mechanisms working side by side, OpenAI's provenance guide describes Content Credentials and SynthID as separate signals in supported output.
A generic metadata inspection cannot establish the presence or absence of a signal in the pixels. If a platform or provider supplies a detector for the signal you are investigating, use it on the file you actually have and interpret the result within that detector's scope. A visual review, an embedded-record check, and a watermark detector each cover different evidence.
Compare the Exports From a Multi-Tool Workflow
Imagine a designer generates a rough concept, retouches part of the picture in another application, and exports a JPEG from a layout tool. The first file might carry a generation prompt. The retouched file might contain a new editing record, carry some earlier fields, or omit them. The final JPEG could still include some of that information and show a different software label. Inspect the three files to see what actually happened.
AI image metadata can survive an earlier stage of a workflow even when a later tool changes the picture. Open the original generator output and the final export separately. Compare the findings by field and value. If you see a tool name in the final file, ask whether it describes generation, later editing, or simply the last application that saved the image. The label alone rarely tells you which pixels were changed.
You can keep a private project folder with a generation export, a working file, and a delivery copy. Use filenames or a project log that tell you which version is which. If a collaborator needs reproducibility, share the relevant settings or references through the project channel. That gives you more control over who sees working details and avoids relying on a delivery image to carry your entire workflow.
For a worked comparison, see Can an AI Image Carry Traces From More Than One AI Tool?. Use that example as a model for a before-and-after check on your own exports.
Inspect an Image Before You Share It
Start with the exact file you plan to attach or upload. If you are posting a JPEG exported from a PNG, inspect the JPEG. If the file goes through a publishing system that produces a new downloadable copy, inspect that copy when you have access to it. Earlier results describe the earlier file.
- Look at the image. Read captions and visible labels. Check backgrounds, screenshots, and text within the picture for details that should stay private.
- Inspect the file. Open it in the No Trace Share workspace and review supported findings, including AI-related fields and detected Content Credentials records.
- Decide what should travel with it. Preserve a source copy when the credential or workflow matters. Make the supported keep or remove choices that fit the purpose of the delivery.
- Clean and review the new copy. Compare the before and after results. Look for the specific prompt, software entry, or other finding you meant to remove. Read any warning about data that remains.
- Use the right verification tool. Use a credential verifier for signature and provenance checks. Use an appropriate detector when you need to check a particular invisible watermark.
No Trace Share runs the supported file inspection and cleaning in your browser and creates a separate copy. The AI metadata guide and coverage page explain which fields and formats the app can check. Keep any generated report with your private files if it includes original values you do not intend to share.
Read the Result Within the Scope of the Check
If the cleaner finds a prompt in a supported field, you have a concrete item to review. If it finds a Content Credential, you can use a verifier to learn more about the recorded history. If it finds no supported AI-related fields, that result describes the embedded fields and records it checks. An image may have passed through a system that writes different data or may have been exported without its earlier records.
Ask what decision you need to make. If the concern is a confidential product name, check the values and visible content for that name. If the concern is whether a signer attested to a particular edit, check the credential. If the concern is whether the image includes a specific pixel signal, use the detector for that signal. Those questions can all come up for the same file and require different evidence.
For a client delivery, keep the original when its history matters, share the copy you reviewed, and provide any required attribution or disclosure through the agreed channel. A deliberate check of the exact file is more informative than guessing from how the artwork looks.