Detection Methodology
How AnyImageDetector analyzes images, converts signals into verdicts, protects uploaded files, and explains uncertainty.
Last updated: 2026-07-20
What the detector reports
AnyImageDetector estimates whether an image is likely to have been generated by an AI system. A result contains three parts:
- AI score — a value from 0 to 1 representing the detected likelihood of AI generation.
- Verdict —
likely AI,uncertain, orlikely human. - Confidence — a plain-language indication of how far the score is from the uncertain range.
The result is an estimate, not proof of authorship, copyright ownership, or intent. It should be combined with provenance, context, reverse-image search, metadata, and human review when a decision carries real consequences.
Processing workflow
- The service validates the file type, size, or public image URL.
- Image content is transmitted over an encrypted connection to dedicated processing infrastructure that performs the requested analysis.
- The returned AI score is normalized into the verdict and confidence labels shown by the web tool and API.
- The application does not add the submitted image file to detection history. Signed-in history contains result data such as the score and verdict, not a copy of the image.
Operational service providers are bound to process data only as needed to deliver and secure the service. Infrastructure and credential details are not exposed publicly because doing so would increase abuse and security risk.
Verdict thresholds
| AI score | Verdict | Confidence |
|---|---|---|
| 0.75–1.00 | Likely AI | High |
| 0.55–0.74 | Likely AI | Medium |
| 0.46–0.54 | Uncertain | Low |
| 0.26–0.45 | Likely human | Medium |
| 0.00–0.25 | Likely human | High |
These thresholds make the output easier to interpret; they do not turn a probabilistic result into certainty.
Signals and model coverage
Detection systems may use statistical patterns involving texture, noise, frequency information, edges, and other image-level artifacts. Performance varies by generator, image style, resolution, and post-processing. A model name shown on the website describes a relevant use case, not a guarantee that every image from that model will be identified.
Product updates
We review detection behavior, error reports, and changes in widely used image generators. Material changes to scoring, thresholds, supported inputs, or this methodology will be dated on this page. See accuracy and limitations before using a result in a high-risk workflow.