AiMetaOff / GUIDES

Check ChatGPT image metadata

Learn what ChatGPT image metadata can reveal, how to inspect Content Credentials, and how to check for a SynthID watermark separately.

Start with the file ChatGPT saved. It can carry Content Credentials that record which software made a claim about the image and how it was created or edited. OpenAI also documents SynthID, an invisible watermark carried in the image itself. The two signals require different checks, especially if the file came from a screenshot, an editor, or a social media post. OpenAI provenance documentation

Where did this image file come from?

Direct ChatGPT download

Open the generated image in ChatGPT and select Save. Keep that file as your reference before opening it in an editor or sharing it. If you received the image from someone else, ask how they obtained it. A PNG extension alone tells you little about the export path. Images in ChatGPT

Screenshot or resaved copy

A screenshot captures the displayed pixels in a new file. It does not carry over the source file’s embedded manifest through that capture. An editor or sharing app may preserve metadata, discard it, or write its own records when it exports a copy. If an old manifest survives a change to the bytes it covers, the file binding can fail.

We compared a historical public ChatGPT-generated PNG from the Content Authenticity Initiative with a Chromium screenshot of that image. The public PNG held one embedded C2PA manifest store. The screenshot had none. The sample’s published history includes a format conversion, so this comparison illustrates what happens to credentials when taking a screenshot of a known file. Fresh direct ChatGPT downloads and API exports were not independently tested. Public sample and manifest report

OpenAI API export

For an Image API output, save the image bytes returned by the API. The documented response carries base64-encoded image data, with PNG as the default and JPEG and WebP as options. Decode the base64 and write the bytes directly before converting or processing the image. OpenAI documents Content Credentials and SynthID for supported image generation across ChatGPT and the API and lists exceptions for some API models nearing deprecation. Compare actual files before assuming that the API and ChatGPT downloads use the same fields. OpenAI Image API guide · OpenAI provenance coverage

What ChatGPT Content Credentials can reveal

C2PA groups provenance records into a manifest containing assertions, a claim that references those assertions, and a digital signature over the claim. An assertion can record a creation or editing action, the software involved, or a source type. A file can contain multiple manifests, including records for earlier stages. The names and actions you see depend on what was recorded and what the reader can display. C2PA Technical Specification

In the public sample, AiMetaOff’s detailed report listed ChatGPT as generator information and showed an action that opened an earlier asset. The companion manifest report lists a creation action in that earlier record naming GPT-4o as the software agent and declaring a trained algorithm as the source. Those fields describe this historical example. Published sample manifest

When you review the details, read the validation results individually. The claim signature check asks whether the signature verifies. The file binding check compares cryptographic hashes against the covered parts of the file. Signer trust depends on whether the validator accepts the signer’s certificate chain under its trust policy. In the public example, AiMetaOff reported verified assertion hashes and file binding, an unverified claim signature, failed certificate validity and unverified signer trust. Treat the software names as recorded declarations. C2PA Technical Specification

Can you find the original ChatGPT prompt?

OpenAI’s provenance documentation describes origin and history fields and makes no promise that the original chat prompt is embedded. We found no original prompt in the inspected records or the published manifest report for the sample. Someone could add text to a file later, and a prompt or description found there should be interpreted in the context of the field that contains it. If you need the exact words used, go back to the conversation or ask the creator. OpenAI provenance documentation

How to check ChatGPT image metadata

Open AiMetaOff’s Inspect File tool and select Choose files. Add the saved image and wait for Ready to review. AiMetaOff reads supported file metadata and runs its embedded Content Credentials check in your browser.

Open View technical details and check the Input findings for C2PA and ordinary metadata such as EXIF, XMP and PNG text. In the Content Credentials report, look for Declared source and history and Verification results. Check the detailed generator information when the overview does not display a generator name. Read Not found as absence within the structures inspected. An unsupported or incomplete check leaves the result unknown.

An embedded credential can also be inspected with a C2PA validator that reports its trust policy. Preserve the original file so you can compare results across tools. C2PA Technical Specification

How to remove supported embedded metadata

If you need a separate copy without a supported embedded C2PA container, keep the original and open the file in AiMetaOff. Under Cleanup scope, clear the other categories and leave C2PA selected. Select the file, click Clean selected, then use Download new copy when the cleanup completes. AiMetaOff supports this operation for recognized C2PA containers in supported PNG, JPEG and WebP files. See supported embedded C2PA cleanup

Open the downloaded copy in Inspect File and check its own Input findings. In our comparison, the cleaned PNG no longer showed embedded C2PA. A separate PNG check confirmed that only the C2PA chunk was removed and that the compressed image data and decoded pixels were unchanged. If AiMetaOff reports limited cleanup or an unsupported structure, use that result when deciding what you can confirm about the file.

C2PA versus SynthID

C2PA manifests describe provenance in structured records. SynthID encodes an invisible watermark in the image content. It is designed to withstand some transformations, including compression and cropping, and can sometimes survive screenshots. A metadata-only cleanup leaves the pixel data in place. AiMetaOff’s local report offers no SynthID check. OpenAI’s provenance overview

For a separate check, OpenAI’s verification tool accepts an uploaded file and checks for supported OpenAI C2PA and SynthID signals. Use an original export when you can. A result with no supported signal leaves the origin undetermined.

What happens to AI labels on social media?

A platform’s AI label reflects checks or disclosures made on that platform. Meta says its AI info labels can draw on industry-standard indicators or on a disclosure made during upload, including on Instagram. Removing an embedded credential from your copy provides no guarantee about the label on a post. Follow the platform’s disclosure requirements for the content you share. Meta’s AI labeling policy

What local verification can establish

Recheck the exact file you plan to share. For supported structures, the report can show which records were found in that copy and which checks succeeded. Credentials can also be stored externally. AiMetaOff skips remote manifest lookups, so an empty local report leaves remote provenance and the image’s origin open. C2PA Technical Specification

FAQ

Can EXIF show whether an image came from ChatGPT?

EXIF may carry descriptive fields, which can be edited. The OpenAI provenance signals described here are C2PA Content Credentials and SynthID. Check Content Credentials separately even when the EXIF section is empty.

Can a C2PA credential tell me whether ChatGPT generated or edited the image?

If the history is recorded and available, an action can distinguish creation from editing or conversion. Review the action and its source declaration together with the validation results. A generator name by itself gives limited information about what happened.

Does removing metadata prove human authorship?

An image’s authorship depends on how it was made. A cleanup report describes changes to the embedded records in a copy of the file.

Inspect your ChatGPT download