Guide
ChatGPT image metadata
An image that comes out of ChatGPT is often more than the picture. Alongside the pixels, the file can carry provenance and workflow metadata that describes how it was made. This page explains what is actually attached and how to produce a clean copy you can verify — without uploading the image anywhere.
How to use it
- 1Save the ChatGPT image to your device.
- 2Drop it into MetaRase and read the scan report; look for C2PA, XMP, IPTC, or AI-workflow indicators.
- 3Choose an output format — PNG for lossless, or JPEG and WebP for smaller files — then clean the image.
- 4Compare the before and after reports to confirm the supported indicators are gone.
- 5Download the rebuilt file, and keep the original if you need its provenance record.
What ChatGPT images actually carry
Exports are not all identical, but the files that come out of OpenAI tools tend to fall into a handful of categories. Not every image has all of them, and a plain screenshot of an image usually carries none.
- C2PA / Content Credentials: a provenance container that can name the tool and note that generative AI was involved.
- XMP fields: editing and export information, occasionally including the model or workflow.
- IPTC fields: sometimes used for a digital-source-type hint.
- Workflow text: some pipelines embed prompt or generation settings inside a metadata block.
- Container and colour data: ICC profiles and format structures that are unrelated to AI but still travel with the file.
What an AI label is really reading
When a social platform shows an AI note, it is usually reading a machine-readable field from the file you uploaded, not guessing from how the image looks. If the export keeps a Content Credentials container, the platform can surface it. Remove that container and, for the supported indicators, there is nothing left for it to read.
Why rebuilding beats editing tags
It is tempting to delete one or two tags and call it done. In practice, metadata lives in several structures at once — EXIF, XMP, IPTC, and provenance containers — and tag editors often leave leftovers or malformed segments behind. Rebuilding the image from decoded pixels produces a fresh, predictable container, and the export can be scanned again to prove what changed.
Where each format keeps its metadata
The format decides where the metadata hides and, in turn, what a tool has to remove.
| Format | Where metadata lives | Notes |
|---|---|---|
| JPEG | APP1 (EXIF/XMP), APP11 (C2PA), APP13 (IPTC) | The most common export from AI tools. |
| PNG | eXIf, iTXt text, caBX (C2PA) | Lossless; keeps quality when re-encoded. |
| WebP | EXIF, XMP, ICCP, C2PA chunks | Small files, well suited to the web. |
| AVIF / HEIC | ISO-BMFF boxes (Exif, XMP, C2PA) | Browser decoding support varies. |
| GIF | Application and comment extensions | First frame only in MetaRase. |
Platform behaviour at a glance
Removing supported metadata changes what the file advertises. It does not change a platform's independent classifier, and behaviour differs by app, version, and region.
| Platform | Reads AI metadata from the file | Notes |
|---|---|---|
| Often | Can show an AI info note sourced from Content Credentials. | |
| X (Twitter) | Often | A “Made with AI” note can come from file metadata or a manual setting. |
| Sometimes | Behaviour varies by app and region. | |
| Sometimes | Provenance may appear on some posts. | |
| Rarely | Usually strips metadata when the image is sent. | |
| Google Photos / Apple Photos | Viewer only | Show metadata in the app; sharing behaviour varies. |
Will removing metadata remove the AI label everywhere?
Not always, and it helps to separate two things. Cleaning a file removes the supported embedded indicators, so a label that was driven by metadata loses its source. A platform that classifies images with its own model, or a visible AI disclaimer drawn into the picture, is unaffected. MetaRase is deliberately honest about that boundary rather than promising a blanket result.
What MetaRase cannot remove
Being precise about the limits matters more than a broad claim.
- Invisible pixel-level watermarks such as SynthID, which are part of the image data.
- Visible AI text, badges, or logos drawn into the picture itself.
- A platform's own AI classifier or moderation system.
- Metadata that a third party re-adds after you export the file.
How to check the result
After cleaning, MetaRase scans the rebuilt file again and shows a before and after report. If the export still lists a supported indicator, review it before sharing; if it is clean, you have a file that no longer advertises the tool it came from.
Frequently asked questions
Does every ChatGPT image have metadata?
Not necessarily. It depends on the tool and how the file was exported. Scanning the actual file is the only reliable way to know what is inside it.
What are C2PA Content Credentials on an AI image?
C2PA, also branded Content Credentials, is a provenance format that can record the tool used and whether generative AI was involved. Platforms read it to display an AI label.
Will removing metadata lower the image quality?
Only through the ordinary effect of re-encoding. Choose PNG for lossless output, or a high JPEG or WebP quality setting to stay close to the original.
Does MetaRase upload my image?
No. Scanning, cleaning, and verification run in your browser, and the image bytes never reach MetaRase servers.
Can it remove SynthID or invisible watermarks?
No. Pixel-level watermarks are part of the image data, not file metadata, and a browser tool cannot reliably remove them.
Is it legal to remove AI metadata?
For images you created, it is generally allowed. Do not strip copyright or ownership information that belongs to someone else.
Why is the cleaned file a different size?
Rebuilding from pixels re-encodes the image, so the byte size changes even when the picture looks the same.
Will this stop Instagram or X from labelling my post?
It removes the supported file-based signals that drive metadata labels. A label you set manually, or a platform's own detection, is separate.
Sources
Inspect, clean, and verify in your browser
Your image bytes never leave your device. MetaRase rebuilds the image locally, then scans the output again before you download it.
Open the cleanerRelated pages
Results describe supported embedded indicators only and cannot guarantee how a platform will classify an image.