Benchmark Results

Tested on a dataset of known AI-generated images and verified real photographs.

92%
AI Images Caught
Out of 200 known AI-generated images
8%
False Positive Rate
Real photos incorrectly flagged as AI
75%
Modern Models
Midjourney v6, DALL-E 3, Flux (the hardest cases)
80%
Text Detection
AI-written content correctly identified

These numbers will change as AI models improve. We update this page when we run new tests. Last updated: June 2, 2026

Tested against: Midjourney v6, DALL-E 3, Stable Diffusion XL, Flux, real camera photographs from various devices

How We Reach a Verdict

Three independent signals are combined into a single weighted confidence score.

55% weight

Visual Analysis

Claude AI examines the image for generation artifacts, unnatural patterns, diffusion model signatures, and logical inconsistencies. Contributes 55% of the combined score.

25% weight

Metadata (EXIF)

Checks for missing camera data, suspicious software markers, and date inconsistencies. Abstains entirely if no metadata is present, with no penalty for clean images. Contributes 25% of the combined score.

20% weight

Error Level Analysis

Re-compresses the image and measures where pixels changed. Edited or synthetically generated regions show abnormal error patterns. Contributes 20% of the combined score.

Where It Struggles

No detector is perfect. Here’s where AIVerify is most likely to have trouble, so you can weigh results accordingly.

Try it yourself

Free trial included. No credit card required.

Download for Windows