Meta labels AI images from C2PA and IPTC metadata: what stripping does
Meta reads C2PA and IPTC invisible metadata to label AI images. What the February 2024 post says, what it leaves open for video, and how to check files.
Meta says it reads C2PA (version 1.3) and IPTC invisible metadata to label images made with tools from Google, OpenAI, Microsoft, Adobe, Midjourney and Shutterstock, and it labels its own output "Imagined with AI". If a file has had that metadata removed, Meta's metadata route has nothing to read, so the poster's own disclosure is what remains. This applies to the February 2024 announcement; check Meta's current pages for newer scope.
What the announcement says
Sume is not on the named list. The announcement describes industry signals, not a Sume integration, so do not assume a Sume image will be labelled automatically.
| Item | Detail |
|---|---|
| Signals read | C2PA (v1.3) and IPTC invisible metadata |
| Named sources | Google, OpenAI, Microsoft, Adobe, Midjourney, Shutterstock |
| Label on Meta AI images | "Imagined with AI" |
| Video and audio | Initially required manual disclosure |
What Sume's docs say about metadata
The Image API docs describe hosted URLs, output_format (png, jpeg, webp, svg) and per-model parameters. They do not describe C2PA or IPTC signing, so this post makes no claim that Sume output carries either. If you need credentials, treat that as something to confirm for your account, and plan the disclosure yourself in the meantime.
Metadata is also easy to lose. A screenshot, a re-encode or a platform's own processing can drop it. Treat embedded signals as a bonus, not as your label.
Check the file you actually publish
Test the file as it leaves your pipeline, not as it left the generator. Download the image, run it through a C2PA or IPTC reader, and record the result next to the job ID. For video, video inspect returns probe facts, stills and optional speech-to-text; the docs list no credential read, so use a separate tool for that.
asset,job_id,after_resize,c2pa_found,iptc_found,manual_label_needed
banner-4x5.jpg,<job id>,yes,no,no,yesA practical rule
The ground is moving. Meta's later statements on ad labels and the EU code are covered in related posts.
- Label manually whenever you cannot show that metadata survived to the posted file.
- Keep the original generator output, untouched, in storage.
- Never strip metadata on purpose to dodge a label; that undermines disclosure.
- Re-read Meta's page for video and audio, which the 2024 post said started as manual.
Where metadata gets lost
Common places to lose embedded metadata are a screenshot, a social platform's own re-compression, an image editor's save-for-web option and a conversion to another format. None of these is malicious, and all of them change what a reader finds.
If you export through several tools, test each export setting once. Some tools keep metadata by default, some remove it by default and some offer a checkbox. Write the working settings into your team's export guide.
For Sume outputs, the safe procedure is to keep the file exactly as downloaded, record the job ID, and add your own disclosure at upload.
Sources
Related posts
More in Use cases
- Meta signs the EU AI content transparency code (July 28, 2026)
Meta confirmed on July 28, 2026 that it signed the EU Code of Practice on AI content transparency. What the date means for teams making ad video.
- Midjourney V8.2 edit stills, animated on Sume
Finish a still in Midjourney V8.2, host it at a public HTTPS URL, and use it as the first frame of a Sume video request. Midjourney itself has no video id here.
- H3 mixing camera, character and audio from three inputs
MiniMax H3 can mix camera movement from one video, a character from an image and audio from another. Sume's minimax-h3 takes image, video and audio references.
- Mixed-language explainer: two TTS jobs joined into one track
A voice that handles 23 or 90 languages still needs one language per request. Render each language as its own job, then join the takes with timeline audio.
Written by Sume