Creators often ask one question after using a metadata remover: “I cleaned the file — why is AI Info still there?”
Sometimes the answer is simple: the post was uploaded before cleaning, or one carousel slide still had C2PA. Other times the trigger was never metadata at all — it was a pixel watermark or visual classifier.
This guide separates metadata labels from pixel-level signals like SynthID, so you can plan honest upload workflows and set realistic expectations.
Start here:
- AI metadata checker — see C2PA, XMP, EXIF, PNG chunks locally
- AI metadata remover — strip file-level markers before upload
- Disclaimer — limits of our tools
Two different “hidden” layers in a file
Think of a raster image as two layers platforms may inspect:
| Layer | Examples | Typical tools that read it | Removed by metadata strip? |
|---|---|---|---|
| File metadata | EXIF GPS, XMP prompts, C2PA manifest, PNG workflow JSON | Instagram/Meta ingest, Pinterest, TikTok upload parsers | Yes (when tool targets those blocks) |
| Pixel signal | SynthID, some vendor-specific watermarks, steganographic marks | Research systems; platform classifiers (varies by product) | No — would alter pixels |
Our site focuses on the first layer — metadata hygiene for JPG, PNG, and WebP you own. We do not claim to modify pixel watermarks.
What is SynthID?
SynthID is a watermarking approach associated with Google DeepMind / Google Cloud, designed to survive common edits (crop, compression) by embedding information in the pixel domain. It is discussed in the context of:
- Imagen and other Google AI image pipelines
- Some experimental detection and identification workflows
SynthID is not stored as a normal EXIF tag you can toggle in Lightroom. A metadata checker showing “no C2PA” does not prove “no SynthID.”
Creator takeaway: If your pipeline is Google-native AI generation, assume non-metadata detection may exist even after a spotless metadata report.
Metadata-driven labels — the common case today
For many Instagram, Facebook, Pinterest, and TikTok still uploads in 2025–2026, labels still trace back to C2PA and XMP:
- DALL·E / ChatGPT exports → C2PA
- Adobe Firefly & Generative Fill → C2PA + IPTC/XMP
- Midjourney saves → XMP parameters
- Canva AI backgrounds → flattened JPG with provenance markers
When those markers are present at upload, AI Info or GenAI flags are expected platform behavior, not a bug.
Fix path when metadata is the trigger:
- Checker confirms C2PA/XMP
- Remover strips targeted blocks locally
- Re-upload a new post with the cleaned file
Walkthrough: remove AI info from photos
When cleaning metadata is not enough
Suspect non-metadata detection if:
- Checker reports no C2PA, no AI XMP, no PNG AI chunks — but label persists on brand-new upload
- Source is a known pixel-watermarked generator
- Platform added label without obvious metadata in your export (rare but reported during policy experiments)
- Video or Reels containers carry provenance separately from a cleaned cover JPEG
What to do:
- Do not assume “broken tool” — compare before/after checker output
- Try a re-export from original RAW or layered project when possible
- Read platform help centers for manual AI disclosure toggles (where offered)
- Review Disclaimer — we do not guarantee platform outcomes
Related: AI label false positives (when metadata was the cause on real photos)
Visual classifiers — the third category
Platforms may run machine-learning models on pixels to estimate synthetic content. These models:
- Change over time with retraining
- Produce false positives and false negatives
- Are largely opaque to creators (no public “score” per upload)
Metadata removers do not interact with classifiers. Caption honesty, platform disclosure tools, and content authenticity policies matter here.
Overview: how platforms detect AI images
Honest workflow diagram
Export final image
↓
Local metadata check
↓
C2PA/XMP present? ──No──→ Label still on new upload? → suspect pixel/visual detection
│
Yes
↓
Local metadata clean
↓
Re-upload cleaned file
↓
Label gone? ──Yes──→ Was metadata-driven (for this upload)
│
No
↓
Document checker report; consider non-metadata causes; follow platform + legal disclosure rulesEthics and platform terms
Stripping metadata to fix accidental false positives on hybrid photography is a common legitimate use. Stripping metadata to deceive viewers, customers, or regulators is not.
The EU AI Act, FTC-style advertising rules, and platform Terms may require visible disclosure even when metadata is empty. See EU AI Act & social images guide.
Related tools & guides
We build browser-local inspection and cleaning tools so you can verify what changed in the file itself — not pixel watermarks — before every upload.
よくある質問
What is SynthID?
SynthID is Google’s imperceptible watermark embedded in pixel data of some AI-generated or AI-edited images. It is not the same as C2PA or EXIF metadata and cannot be removed by metadata-only tools.
Can metadata removers delete SynthID?
No. Metadata removers strip EXIF, XMP, C2PA, and PNG text chunks. SynthID lives in the image signal itself; removing it would require changing pixels.
Why does Instagram show AI Info if metadata is clean?
Possible reasons include pixel-level watermarks, visual classifiers, prior platform-side analysis, or labels applied before you re-uploaded a cleaned file. Metadata cleaning only addresses file-tag triggers.
Should I still clean C2PA before upload?
Yes when your goal is to avoid metadata-driven labels. Keep expectations realistic if the platform also uses non-metadata detection.