AI Image Metadata: AIGC Labels, Signatures and EXIF
Metadata is the most reliable evidence that an image is AI-generated — if it survived. This guide covers what AI tools actually write into files, what China's mandatory standard requires, and which everyday operations erase each signal.
Metadata is structured information embedded in a file that does not affect how the image displays. The shooting parameters in a JPG's properties panel, a text chunk hidden inside a PNG, the generator name written on export — all metadata.
The distinction that matters is where it lives:
- Container metadata (EXIF, XMP, C2PA manifests) — stored in the file structure. Re-wrapping the file loses it. A screenshot always loses it.
- Pixel-embedded data (SynthID watermarks) — stored in the pixel values. A screenshot copies them along with the pixels.
1. What AI tools actually write into files
The fields differ by tool, but the idea is consistent: record who, when, with what, and with which parameters.
| Tool | Common location | Typical fields |
|---|---|---|
| Midjourney | PNG tEXt / JPEG XMP | Software, prompt, parameter string, version |
| Stable Diffusion / A1111 | PNG tEXt | parameters (sampler, steps, CFG, seed, prompt) |
| ComfyUI | PNG tEXt | prompt, workflow (full node-graph JSON) |
| DALL·E / ChatGPT | PNG tEXt or XMP | software name, prompt, date |
| NovelAI | PNG tEXt | Source, Description, Software |
| Flux.1 | PNG tEXt | generator identifier and parameters |
| Camera capture | EXIF | Make, Model, Lens, aperture, shutter, ISO, time, GPS |
One practical detail: ComfyUI writes the entire node graph into the PNG, including model filenames, prompts and even LoRA names. Useful for attribution, but it is also how prompts leak — worth checking before you share an image.
2. China's implicit AIGC label (GB 45438-2025)
China runs the world's first mandatory national standard for AI content labelling. The 《Measures for Labelling AI-Generated Synthetic Content》 and the compulsory national standard GB 45438-2025 have been in force since 1 September 2025, after a six-month transition period.
The standard defines two forms of labelling:
Explicit labels
Visible notices: an "AI generated" text badge overlaid on the image, an audible cue in audio or video. The goal is that an ordinary person sees it at a glance.
Implicit labels
Machine-readable markers written into file metadata. The requirements:
- Field names must contain an AIGC marker — a hard requirement that lets detection tools find them by name.
- Record AI generation confirmation status (yes / no / suspected).
- Record the generating service provider and its code.
- Record the content distribution platform.
- Record a unique identifier.
- Provide a digital signature or hash for tamper resistance.
The corresponding XMP namespace is http://www.tc260.org.cn/ns/AIGC/1.0/. A companion document, TC260-PG-20252A, specifies how provider codes are constructed.
The standard also prohibits removal
The Measures explicitly forbid any organisation or individual from maliciously deleting, altering, forging or concealing AI-generated content labels, and from providing tools or services to do so for others. In other words, stripping the label is itself a regulated act — which is what makes metadata detection meaningful in compliance contexts.
3. C2PA Content Credentials: the international manifest
Content Credentials, defined by the Coalition for Content Provenance and Authenticity, are a digitally signed provenance manifest recording who produced the file, with which tool, and which tools touched it afterwards. Because it is signed, any modification invalidates the signature.
It is an open standard, published at spec.c2pa.org, with verification tools including the C2PA Viewer and verify.contentauthenticity.org. Adobe Firefly and the ChatGPT image tool are among those writing C2PA manifests.
Its fatal weakness is the mirror image of its strength: the manifest lives in the file container, so signing guarantees "not tampered with" but also means one screenshot erases it entirely. Platform transcoding usually strips container metadata as well.
4. SynthID: the watermark hidden in pixels
SynthID is Google DeepMind's invisible watermarking scheme. It distributes the marker imperceptibly across the pixel values of the whole image. Because the information lives in the pixels, a screenshot copies it along, and it survives JPEG compression, resizing and moderate cropping.
What it can express is limited: it says "this came from a SynthID-enabled generator" and nothing about creator, timestamp or edit history. Reading it requires the matching detector — a generic AI image detection tool cannot read it.
In May 2026 OpenAI joined the C2PA steering committee and announced it would embed both C2PA manifests and SynthID watermarks in images from ChatGPT and its API; Google announced that C2PA and SynthID detection would come natively to Google Search and Chrome. The dual-layer approach is becoming the de facto standard — manifests for complete provenance, watermarks as a fallback once the manifest is gone.
5. Camera EXIF: evidence of a real photograph
Camera EXIF runs on the opposite logic to AI labelling. A photo carrying a complete camera record is strong evidence of real capture: Make, Model and Lens, plus aperture, shutter, ISO, focal length, capture time and often GPS coordinates.
Generators do not fabricate these. So while missing metadata is evidence in neither direction, complete metadata pointing at a real camera is a genuinely reliable positive signal.
6. Which operations erase metadata
| Operation | Container metadata (EXIF/XMP/C2PA) | Pixel watermark (SynthID) |
|---|---|---|
| Screenshot | Completely lost | Preserved |
| Re-save as JPEG (standard tools) | Usually lost | Mostly preserved |
| Social platform upload | Usually lost | Platform-dependent |
| Image compression (e.g. tinypng) | Usually lost | Partially degraded |
| Resize / crop | Usually lost | Mostly preserved |
| Sending via WeChat | Usually lost | Version-dependent |
The pattern is clear: any operation that writes a new file strips container metadata; operations that only alter pixel values preserve pixel watermarks.
Privacy note
Camera EXIF can contain GPS coordinates and timestamps precise to the second. If you would rather not reveal your location, checking and cleaning it before sharing is worth doing — that is privacy hygiene, not concealment.
7. Reading metadata yourself
No software installation is required for a first look:
- macOS: right-click → Open With → Preview, then ⌘I; or Get Info.
- Windows: right-click → Properties → Details; EXIF needs an extra tool.
- Command line:
exiftool photo.pngis the most complete option and is cross-platform. Without installing anything,PIL.Image.open(f).infoin Python shows most PNG text chunks. - Online: verify.contentauthenticity.org checks C2PA credentials entirely in your browser.
To see "metadata clues" and "pixel analysis" as two independent results at once, upload the image here. The tool parses PNG and JPEG metadata, identifies TC260 AIGC markers and interprets them semantically, matches known generator signatures, reads camera EXIF, and presents the result alongside the pixel-level model analysis.
Frequently asked questions
How do I check if an image contains AI generation info?
On macOS open it in Preview and press ⌘I; on Windows check Properties → Details; for the full picture use exiftool on the command line. Generator info usually sits in a PNG tEXt chunk, and for JPEGs in the XMP block or the EXIF Software field. Uploading to this site lists the same findings.
Are AIGC labels and C2PA the same thing?
No. AIGC labelling is China's GB 45438-2025 system: field names must contain an AIGC marker, and the metadata records generation status, provider, platform, a unique ID and a signature or hash. C2PA is the international standard, where Content Credentials are a digitally signed provenance manifest recording creator, tool, time and edit history. The two can coexist in one file.
Can you still detect AI traces after a screenshot?
It depends on the watermark. Container metadata (EXIF, XMP, C2PA manifests) disappears entirely in a screenshot. Pixel watermarks such as SynthID are copied along with the pixels, so detection may still succeed. With neither present, only pixel statistics remain, and accuracy drops noticeably.
Can images received through WeChat still be checked?
Usually not well. WeChat re-encodes images in transit, which strips container metadata. Pixel watermarks like SynthID are a separate matter. For any image that has already been compressed by a platform, treat a detection result as inference rather than confirmation.
Is removing metadata illegal?
In China, the 《Measures for Labelling AI-Generated Synthetic Content》 explicitly prohibit maliciously deleting, altering, forging or concealing AI-generated content labels, and prohibit providing tools or services for others to do so. Note the distinction: the rule targets malicious removal of AI labels — it is not the same as clearing GPS coordinates for privacy reasons.
Related articles
- Why AI Image Detectors Fail (And What to Do)
- How to Tell If an Image Is AI-Generated: 7 Methods
- AI Image Detection: Metadata vs C2PA vs Watermarks
- Midjourney vs Real Photos: How to Tell Them Apart
Check an image yourself
Upload an image to see metadata clues and pixel analysis together
Open the detector