AI image metadata
SynthID and Invisible Watermarks: What Metadata Cleaning Cannot Remove
Published November 8, 2026 · 7 min read
Pixel-level watermarks like Google's SynthID survive metadata removal, re-encoding and screenshots. Understand what they are, what survives, and what that means for AI images.
Two completely different hiding places
AI-related markings live in two places that behave very differently. Metadata-based markers — XMP fields, PNG text chunks, C2PA manifests — are data stored beside the pixels. Pixel-based watermarks — SynthID among them — are patterns embedded in the pixels themselves, woven in by the generator so the image looks identical to humans but reads as AI-generated to a detector.
What survives what
- Metadata markers: removed by any proper re-encode or metadata strip.
- Pixel watermarks: survive re-encoding, format conversion, resizing, compression and screenshots, by design.
- Pixel watermarks are detectable only by the corresponding detector, which is usually not public.
Why the honest answer is "no"
If your goal is to make an AI image undetectable as AI, metadata cleaning is not sufficient and never will be. SynthID's own documentation states it survives common edits. Any tool claiming to "remove AI detection" from an image outright is overpromising: detectors can also flag AI images from statistical patterns in the pixels, with error rates in both directions.
This is why the honest framing is about metadata: a cleaner removes the generation records stored in the file — prompts, tool names, provenance manifests — and proves their absence by re-scanning. What it cannot change is what the pixels themselves reveal to a watermark detector.
The privacy angle still matters
There is a second reason to care about metadata that has nothing to do with AI detection: prompts leak. A PNG from a diffusion tool often contains the full prompt, negative prompt, sampler and model version. If a prompt includes personal details, or if you are sharing a client's file and do not want your workflow exposed, metadata removal is exactly the right tool for that job. What Stable Diffusion tools write into PNG files shows where those records live.
The takeaway: clean metadata deliberately and verify the result, and do not rely on any tool — including this one — to defeat pixel watermarks. Those two claims should never be confused.
Check your own files
Inspect and clean metadata from images and PDFs directly in your browser. Files are never uploaded to a server.
Remove an image's metadata →Related guides
- Does Removing EXIF Data Remove AI Labels?Short answer: it depends where the AI label is stored. A field-by-field breakdown of EXIF, XMP, PNG text chunks, C2PA records, and what a full cleaning pass actually removes.
- What Is C2PA? Content Credentials ExplainedC2PA Content Credentials are signed provenance records attached to photos and PDFs. Here is how the standard works, where it is stored, and what removing it does and does not mean.
- How to Check What Metadata Your Files Really ContainYour photos, PDFs and screenshots carry more than you think. A guide to viewing metadata across formats and file types — and what to look for first.