C2PA vs SynthID vs IPTC: The Three Ways AI Images Are Labeled
IPTC tags, C2PA manifests and invisible watermarks label AI images in different ways. Where each lives, what survives a screenshot, and what can be removed.
The people who build BlankAI
In This Article
- 01IPTC: a simple, standard word
- 02C2PA: a signed history
- 03Watermarks: a mark that travels with the pixels
- 04Why platforms combine them
- 05What you can control
When a platform decides an image is AI-generated, it is usually reading one of three things: an IPTC tag, a C2PA manifest, or an invisible watermark. They were designed for different jobs, and they fail in different ways.
| IPTC digital source type | C2PA manifest | Invisible watermark (SynthID etc.) | |
|---|---|---|---|
| Stored in | XMP or IPTC metadata | JUMBF box in the file | The pixels |
| Signed | No | Yes | No, but proprietary |
| Survives re-encoding | No | No | Yes |
| Survives a screenshot | No | No | Designed to |
| Who writes it | Midjourney, Photoshop, Apple, many apps | OpenAI, Google, Adobe, Microsoft, TikTok, cameras | Google, OpenAI, Meta, TikTok, ByteDance |
| Who reads it | Instagram, Pinterest | LinkedIn, TikTok, Meta, verifiers | The vendor's own checker |
| Removed by a metadata cleaner | Yes | Yes | No |
IPTC: a simple, standard word
The IPTC digital source type is a field with a value from a fixed vocabulary. trainedAlgorithmicMedia means an image made by AI; compositeWithTrainedAlgorithmicMedia means a real image with AI-edited parts. It is easy to write and easy to read, which is why platforms adopted it first, and just as easy to lose.
C2PA: a signed history
A C2PA manifest describes what made a file and what was done to it, and it is signed, so tampering is detectable. It can carry the same digital source type inside its actions. But it is still file data: re-encoding or stripping the file removes it, and its absence proves nothing.
Watermarks: a mark that travels with the pixels
Watermarks exist because metadata is fragile. SynthID, Meta's Content Seal and ByteDance's watermarks sit in the pixel values and survive the things that remove metadata. The trade-off is that each vendor mostly detects its own: Meta's detector does not read SynthID, and Google's does not read Content Seal.
Why platforms combine them
The EU Code of Practice on AI-generated content asks providers for at least two layers, signed metadata plus an imperceptible watermark, precisely because each covers the other's weakness. Verifiers are starting to cross-check them: a 2026 paper describes how a watermark that contradicts a file's credentials, or appears where credentials are missing, is itself informative.
What you can control
The first two columns are file metadata, so you can see and remove them. The watermark is not in your control once the image is generated. If a label came from an editing tool's metadata on a genuine photo, cleaning a copy removes that signal. If the image is AI-generated, the watermark and the platforms' disclosure rules still apply.
BlankAI reports IPTC source types and C2PA manifests in any file, and tells you which watermark its generator uses.
Check a fileSources
- IPTC Digital Source Type vocabulary
- C2PA Technical Specification 2.4
- Google: New ways to identify AI-generated media (May 2026)
- Meta AI: Introducing Muse Image and Muse Video
- IPTC: AI disclosure on social media, a work in progress (2026)
- Integrity Clash: when provenance metadata and watermarks disagree (arXiv 2603.02378)
- European Commission: Code of Practice on transparency of AI-generated content
Share this article
BlankAI Team
The people who build BlankAI
BlankAI builds browser-based tools for inspecting and cleaning image metadata. These guides are written by the same team, checked against vendor documentation, standards and independent tests, and updated when those change. Sources are listed at the end of each article. Corrections: [email protected].