AI image metadata
Does Removing EXIF Data Remove AI Labels?
Published October 28, 2026 · 7 min read
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.
"AI labels" are not one thing
When people ask whether removing EXIF removes AI labels, they are usually picturing one switch. In reality, AI-generation information is written by different tools into different structures, and EXIF is only one of them. Here is the field-by-field map.
What EXIF itself can contain
So yes — EXIF can carry AI labels, and stripping EXIF removes those particular fields. But that is not the whole story.
- Software field: names the generating or editing application (for example a Stable Diffusion build or a Photoshop version).
- ImageDescription and UserComment: free-text fields that tools like AUTOMATIC1111 fill with the full generation prompt and settings.
- Artist / Copyright: sometimes set by AI tools or stock exporters.
The structures EXIF removal does not touch
- XMP metadata: xmp:CreatorTool naming Firefly, DALL·E or Midjourney, plus Iptc4xmpExt:DigitalSourceType values like trainedAlgorithmicMedia, and the xmpMM:History edit record.
- PNG text chunks: tEXt/iTXt keys named parameters (AUTOMATIC1111), prompt and workflow (ComfyUI), Comment (NovelAI), and invokeai_metadata.
- C2PA Content Credentials: the APP11 manifest in JPEG, the caBX chunk in PNG, equivalent chunks in WebP and AVIF.
- JPEG comment segments: a standalone FFFE comment marker some pipelines use.
What actually works
A complete pass removes all of these structures at once. The practical way to do that without byte-level surgery is re-encoding: decode the image and write a new file from the decoded pixels in a fresh container. The new file has no inherited XMP, no PNG text chunks, no APP11 segment, no comment segment — nothing survives unless the encoder deliberately adds it.
This is exactly how the free image metadata cleaner works: it re-encodes your file in the browser, then re-scans the output with the same inspector used on the input and shows you what remains, so "clean" is a measurement rather than a promise.
The one thing no metadata tool can remove
Invisible watermarks embedded in the pixels themselves — Google's SynthID and similar schemes — are not metadata. They survive re-encoding, resizing and screenshots. If an image carries a pixel watermark, no metadata cleaner can remove it, and anyone claiming otherwise is wrong. The SynthID guide explains what survives and what does not.
Check your own files
Inspect and clean metadata from images and PDFs directly in your browser. Files are never uploaded to a server.
Scan an image for AI markers →Related guides
- 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.
- SynthID and Invisible Watermarks: What Metadata Cleaning Cannot RemovePixel-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.
- Stable Diffusion Metadata: Parameters, Prompts and Workflows in PNG FilesPNGs from AUTOMATIC1111, ComfyUI, NovelAI and InvokeAI carry the full generation recipe in text chunks. Here is what each tool writes and how to read or remove it.