How accurate are AI image detectors?

An accuracy figure is a benchmark result — performance on a curated test set. The test set determines what the number means. A detector trained on 2023 Stable Diffusion outputs and tested on 2023 Stable Diffusion outputs will report a very high number, but tells you nothing about its performance on a 2026 model you actually need to evaluate.

Real images are compressed, recompressed, filtered, screenshotted, and re-saved before they reach you. Each step strips signal. Detectors that performed well on pristine generator outputs lose ground on Instagram-recompressed copies of the same images.

Real human photos also drift toward 'looks AI' under heavy retouching, professional lighting, and beauty filters — common on social media. False positives on real images are the most common complaint among users.

On clean, original AI images from common generators: 85–95% true-positive rate is realistic. On heavily compressed social-media reposts: closer to 60–80%. On adversarially-tuned outputs designed to evade detectors: anywhere from baseline to barely above chance.

These are rough ranges, not vendor claims — the right takeaway is that any single detection score is a probability, not a verdict.

Treat 'high accuracy' marketing claims as a flag for skepticism, not confidence. The detectors with the most honest numbers are the ones that publish their failure modes. Use detection as one input alongside provenance, source verification, and the visual checks human eyes are still good at — see the linked guide on detecting AI images.