What an accuracy statement should mean
Detector performance depends on the artwork, its file history, the generators represented in evaluation data and the decision threshold. A single percentage without a test set, date, sample makeup and definition of a correct result is difficult to interpret.
For that reason, a checker should explain its signals and acknowledge uncertainty. Sample figures shown elsewhere on this site are illustrative until supported by a published evaluation with a described methodology.
Why a result can be wrong
- Unfamiliar generators: a model may not resemble systems represented in a detector's reference data.
- Editing and export: repainting, upscaling, screenshots, scans and compression can weaken or alter detectable traces.
- Human-made variation: unusual techniques and digital brushes can share visual properties with generated output.
- AI-assisted workflows: a small generative edit may be difficult to distinguish from the rest of a human-made piece.
- Missing provenance: metadata may be removed or describe only one stage in a longer process.
What confidence can tell you
A confidence score describes how strongly a detector's evidence supports a classification under its own method. It is not automatically the probability that a person did or did not make the artwork; that interpretation requires calibration and validation.
Look at the breakdown. If metadata points one way while pixel analysis points another, that disagreement matters. If the file is a compressed repost, confidence may deserve less weight.
Use results as one part of a review
For consequential decisions about an artist, student or seller, do not rely on an automated score alone. Consider source files, sketches, layers, drafts, signed credentials, process records and the specific policy that applies. Give the creator a fair chance to explain the work.
AI Artwork Checker is intended to support investigation. It does not authenticate ownership or replace expert judgment.
Evaluation transparency
Before relying on a detector for systematic screening, ask which generators and art forms were included, how human-made references were collected, whether data was independent of training, and how false positives and negatives vary by medium.
See how the checks work.