Methodology · v0.1

How Our AI Scanner Works

FreeAIScanner is an authenticity triage tool, not a source-of-truth machine. This methodology explains what each evidence layer can support and where it can fail.

Evidence layer 1: provenance and Content Credentials

Signed provenance can record assertions about origin, editing, and the software involved. When a supported signature validates and contains a direct generator declaration, it can be strong evidence about that file. It does not establish the full history of every visual element, and its absence is not evidence of human creation.

Evidence layer 2: file and camera metadata

File dimensions, encoding, software fields, timestamps, and camera metadata can add context. They are easy to remove or alter. Screenshots, messaging apps, social platforms, and editors routinely rewrite files, so metadata is reported as an observation rather than a verdict.

Evidence layer 3: model inference

A visual classifier estimates whether patterns resemble examples in its training and evaluation data. Performance can shift across generators, editing tools, subjects, resolutions, and transformations. Provider output will be normalized and displayed with model/version information only after benchmark and contract review.

How evidence becomes a result state

The decision engine preserves seven outcomes: direct AI provenance found; likely AI-generated or AI-edited; mixed or uncertain; likely camera or human origin; insufficient evidence; unsupported or unreadable; and scan not completed. Stronger language requires stronger, consistent evidence. Provider failure never becomes a partial verdict.

Known limitations and failure modes

Recompression, screenshots, cropping, low resolution, composite images, partial edits, adversarial changes, and unfamiliar generators can weaken or conflict with available signals. Results must not be used alone for discipline, grading, employment, credit, housing, legal, or similar consequential decisions.

Model and methodology changes

Production releases must identify the pipeline version, classifier model, evaluation date, decision bands, and important changes. A reproducible benchmark is required before accuracy claims or a public benchmark route can be published.

Return home