Our current evidence

The candidate on-device model is not enabled. Evaluation on a licensed test set found too many real images flagged as AI-generated, especially older phone photos and screenshots. The evidence checker does not use this model: it reads provenance information locally instead. We make no claim of validated detection accuracy. Other models or data will be evaluated before reconsidering the launch criteria; there is no promised release date.

Calibration before testing

Model candidates are compared using calibration data only. This separate split also sets the decision boundaries. The upper boundary permits a real-image false-positive rate of at most 5% on calibration data. Held-out test images then measure performance without choosing a better-looking model from their results. Launch requires the false-positive criterion and mainstream detection targets: at least 85% for the standard tier, or 70% to below 85% for the widened tier. A failed gate keeps the model disabled.

Three bands and honest denominators

The model method uses Likely AI-generated, Uncertain and Likely authentic bands. These labels are estimates, not authenticity certificates or probabilities. The lower boundary limits AI images called authentic on calibration data; the widened tier uses a stricter budget and a wider Uncertain band. True-positive rate counts AI images given the AI label. False-positive rate counts real images given that label. Uncertain images remain in both denominators. AI-called-authentic and Uncertain shares are reported separately, with exact counts and Wilson 95% intervals. Empty denominators yield no percentage.

Licensed sources and grouped splits

The test corpus uses public sources with documented licences and attribution. Real categories include recent and older phone photos, camera images, social-media images and screenshots. Rendered screenshots contain licensed real photographs; simulated social-media images are identified separately and do not represent measured platform returns. AI-generated images are never used as real examples. Related images stay together when calibration and test groups are separated, using available prompt or author information. Missing prompt information limits independence. Pixel-only evaluation removes metadata, and the same evaluation processing is applied symmetrically to AI and real images. No processing-specific performance or sample images are published here.

Training overlap and coverage limits

Known training-overlap sources are excluded for the selected model: Qwen-Image-Bench, Rapidata Flux-2-pro and Recraft-v3, VISION, and COCO train2017. Public training hashes help exclude exact matches, but near-duplicate training overlap has not been ruled out. Licence records do not establish that a sample represents every device, author or generator version. GPT Image 2/2.5, nano-banana-2 and later Gemini image models, Grok Imagine, and Seedream 5 remain uncovered. Generator groups below the minimum sample count carry no performance claim.

How this page is populated

The public page currently shows the evaluation method and disabled status. Model metrics appear only after the evaluation gate permits launch and the model is enabled in the build. The published summary contains held-out base-image counts, per-generator detection and uncertainty, real-category false positives and uncertainty, model source and licence, boundaries and evaluation date. It contains aggregate data only. Invalid or inconsistent summaries display no metrics. Development fixtures and mock scores never establish real-world accuracy.

What is a false positive in AI image detection?

A false positive is a real image that a detector labels as AI-generated. This can harm people when used to judge work, applications or news photographs. Review provenance, the original context and independent evidence; never make an important decision about a person using a detector label alone.

Why not publish one accuracy number?

Performance depends on generator versions, image categories and sampling. A single average can hide real-image errors and weak generator coverage. Exact denominators, category-level results and uncertainty intervals make an allowed report easier to assess. They still describe its test set rather than proving that every future image will receive a correct label.