Methodology
Img2det ranks evidence by provenance strength and uncertainty.
The checker is designed for file evidence review. It explains what the uploaded file contains, which signal drove the report, and which claims remain outside the file. That is why the product avoids binary real-or-fake promises.
Evidence hierarchy
Layer 1
Trusted C2PA verification
What it can support: Strongest file-level provenance when the manifest, signature, trust policy, and asset binding all validate.
Limit: It does not prove the caption, consent, legality, or real-world truth of the depicted event.
Layer 2
AI provenance markers
What it can support: Useful when provider names, trainedAlgorithmicMedia strings, or C2PA-like markers appear in the original bytes.
Limit: Marker-only evidence is not cryptographic verification and should not be described as trusted provenance.
Layer 3
Camera-like metadata
What it can support: Can support a camera-origin interpretation when plausible Make, Model, and formation clues are present.
Limit: EXIF can be absent, edited, stripped, or copied, so camera-like support is not proof of authenticity.
Layer 4
Byte and frequency clues
What it can support: Helpful for surfacing remnants, compression patterns, and forensic context when stronger provenance is missing.
Limit: These clues are supportive and uncalibrated; they are not legal attribution or standalone AI probabilities.
Layer 5
Source context
What it can support: Necessary for publication decisions because file evidence alone cannot verify every claim around an image.
Limit: Source context lives outside the uploaded file and must be documented by the reviewer.
Why the report is not a binary AI detector
A detector-style score can be affected by compression, resizing, screenshots, model changes, and editing workflows. Img2det keeps those signals in a supporting role and gives priority to provenance evidence that can be inspected directly.
The final assessment is therefore a short evidence summary, not a universal truth claim. A file can contain trusted provenance while the caption remains misleading, or it can be stripped of metadata while still being a legitimate photograph.
How reviewers should use the output
Start with the Final assessment, then inspect the main driver and detailed evidence matrix. If the report says marker-only, preserve that wording. If it says no C2PA data, write that provenance was not found in this file, not that the image is fake.
When a decision affects publication, moderation, or reputation, combine the file report with source checks: original file request, uploader identity, previous publication history, reverse search, archive links, and caption review.
Next step
Use the examples page to see how this methodology turns into practical report language, then run the live checker with the best original file you can obtain.