The evidence scale
The result describes the evidence inside the file, not how convincing the image looks.
Confirmed AI-generated
Credentials or IPTC explicitly declare AI generation.
AI tool traces found
Metadata contains AI tool names or generation parameters.
No AI traces found
Missing evidence is not proof of non-AI.
Declared capture / human creation
Credentials declare capture or human creation; they do not prove a real scene.
Strength of a declaration, not a probability score. Capture credentials are a separate provenance claim. Signature demo only: algorithmicMedia, not a camera capture.
Which fields do we read?
How this checker reads the evidence
01Choose the original file
Drop, select or paste an image. Original downloads are more likely to retain metadata than screenshots or social-media copies.
02Read its provenance
The browser checks embedded C2PA manifests, IPTC/XMP, camera EXIF and PNG generation parameters. Evidence is ranked by source.
03Review the actual evidence
Inspect the source fields, signature status, actions and ingredients. A label describes what the metadata declares, not a guarantee that a scene is real.
| Source | Fields inspected | Strength | Common in |
|---|---|---|---|
| C2PA | Signature, file binding, digitalSourceType, actions, ingredients | ●●●● | Firefly / signed capture |
| IPTC · XMP | DigitalSourceType, CreatorTool, Software | ●●●○ | AI-labelled exports |
| EXIF | Make, Model, Software, UserComment | ●○○○ | Cameras / photo editors |
| PNG text | parameters, prompt, workflow | ●●○○ | SD WebUI / ComfyUI |
| Tool signatures | 20+ tool names and generation parameter patterns | ●●○○ | AI generation tools |
Dots describe typical evidence strength, not authenticity. C2PA strength depends on validation and the declared source; IPTC is self-declared.
A missing label is not proof of a real photo
- This is not a pixel classifier. It cannot spot AI artifacts or read invisible watermarks such as SynthID; those require the relevant official tools.
- Metadata can be removed or forged. Social platforms, screenshots and re-saving often strip it. No credentials does not mean non-AI.
- A valid signature means the SDK validated the embedded claim and its bindings. This tool does not establish signer identity or certificate trust online. An invalid or unverified C2PA claim is only a hint.
- An IPTC AI label is a self-declared source type and is not cryptographically verified. Camera make/model is weak evidence only. AI editing can coexist with a camera capture history.
Questions about AI image detection
Why was my AI image not detected?
Many generators do not embed metadata, and platforms often remove it. If no AI metadata survives, this checker cannot determine the image's origin. It does not analyze pixels.
What are C2PA Content Credentials?
C2PA is a standard for signed provenance records. A manifest can describe software, edits, ingredients and the declared source type. A signature protects that record; it does not prove that every claim or depicted event is true.
Can it detect AI after taking a screenshot?
Usually not from the original metadata. A screenshot creates a new file and normally loses the original credentials. Use the original downloaded image when possible.
Which image formats are supported?
JPEG, PNG and WebP, plus AVIF and HEIC/HEIF when supported by the reader. The limit is 20 MB per file. Browser previews for HEIC may be unavailable even when metadata is readable. Clipboard images often arrive without original metadata.
How accurate is this AI image detector?
It reports evidence, not a probability score. An AI source declaration, an AI tool hint and missing metadata are different outcomes. Metadata can be forged; absence of evidence never proves human authorship. Unsupported or unreadable credentials are reported as incomplete checks.