AI Image Detection Methodology
OriginVerdict is designed as a second-look tool, not an authenticity certificate. This page explains what the current AI photo detector examines, how its labels are assigned, and where the method can fail.
What the detector analyzes
Each supported image is evaluated in two separate ways:
- Pixel-based AI signal. The image is sent to Sightengine's
genaimodel. The model analyzes visual content and returns a score from 0 to 1 indicating how strongly the image resembles AI-generated or AI-edited content. According to Sightengine, this model does not depend on EXIF metadata or watermarks for its score. - Limited file metadata. OriginVerdict independently reads a small set of EXIF fields when present: camera make, camera model, capture time, dimensions, and editing-software name. GPS parsing is disabled.
Metadata does not change the pixel-model score. It is shown as context because a camera tag, an editing-software tag, or missing EXIF can be useful to a reviewer without proving origin.
Score bands
OriginVerdict maps the provider score into three intentionally broad bands:
| AI signal | Label | Interpretation |
|---|---|---|
| 0–20% | Low AI-generation signal | Few visual patterns associated with generative AI were detected. This does not authenticate the source or photographer. |
| 21–79% | Inconclusive | The available signal is not strong enough for a restrained binary conclusion. |
| 80–100% | Likely AI-generated or AI-edited | The image contains patterns associated with generated or AI-edited imagery. This remains a model assessment, not proof. |
These thresholds are product interpretation bands, not universal scientific cutoffs. They are deliberately conservative so that mid-range results remain inconclusive.
Supported files
The current tool accepts JPG, PNG, and WebP images up to 10 MB. It validates the file signature on the server rather than trusting only the filename or browser-provided MIME type.
Known failure modes
False positives and false negatives are possible. Factors that can affect a result include:
- screenshots and photographs of screens;
- heavy compression, resizing, cropping, filters, or repeated re-encoding;
- composites containing both camera-made and generated regions;
- subtle AI edits to an otherwise real photograph;
- unusual camera pipelines, digital art, CGI, or synthetic-looking real scenes;
- generators or editing methods that are new or underrepresented in model training; and
- adversarial attempts to evade detection.
How to use a result
For a consequential review, combine the score with the original source, publication history, reverse-image search, content credentials when available, file provenance, and human review. Do not accuse a person of deception solely because a detector returns a high score.
OriginVerdict does not currently verify C2PA Content Credentials and does not claim to identify the exact generator responsible for an image.
Privacy and processing
The browser uploads the original to private temporary storage. OriginVerdict reads it for the synchronous check and attempts to delete it immediately after the request; a one-day lifecycle rule covers interrupted cleanup. The image is also transmitted to Sightengine for analysis, so Sightengine's processing and retention terms apply. See the Privacy Policy for details.
Changes and evaluation
Detection models and generators change over time. We will revise this page when the provider, score interpretation, supported formats, or result behavior materially changes. Before publishing accuracy claims, OriginVerdict will record the dataset, sample composition, test date, and error definitions used for that evaluation.
Questions or reproducible counterexamples can be sent to support@originverdict.com.