AI Image Detection Statistics: Can People Tell What’s Real?
AI image detection accuracy depends on the images, participants and task. A 2025 Microsoft Research study reports 62% overall accuracy; a portrait-only preprint revised in August 2026 reports 85.2%. These are different experiments, not a time series showing that detection improved by 23.2 points.
A single percentage cannot describe whether people can identify every AI image. Overall accuracy, correctly identifying synthetic images and completing an entire quiz without mistakes use different denominators. This page compares studies within their stated scope and identifies what the results cannot prove.
Try it on Morphed: Explore Morphed.
Key statistics
Microsoft Research’s June 2025 publication reports accuracy across real and AI-generated images. Participants found portraits easier than natural and urban landscapes.
The study analyzed approximately 287,000 image evaluations from global participants in the online Real or Not Quiz. This is not a representative population survey.
The Kim & Kim preprint reports this result for a specific portrait task. Version 2 was posted August 23, 2026; the generators were available in July 2025.
Human AI image detection studies compared
The results below are not pooled. Each study has its own sample, stimuli and outcome definition.
| Study | Result | Sample / task | Scope and limitation |
|---|---|---|---|
| Microsoft Research, June 2025 | 62% overall accuracy | 12,500+ participants; ~287,000 evaluations | Mixed image categories; online quiz |
| Kim & Kim, August 2026 revision | 85.2% mean accuracy | 1,664 participants; 20 trials each | Portraits; FFHQ real images, ChatGPT-4o and Imagen 3 outputs; preprint |
Links to both original research records are in the cards and Sources. The portrait study used a July 2025 model snapshot, not every model available at the September 2026 review date.
What do the detection studies actually show?
The studies support different conclusions under different conditions. Sampling, image selection, generator versions and device context matter; neither result establishes a universal detection rate.
Adults aged 20–69 classified 20 portraits each from a 210-image pool. Fieldwork ran August 2025–January 2026.
How we compiled this data
Reviewed September 22, 2026 against Microsoft Research’s publication summary and version 2 of the portrait preprint. We retained sample sizes, fieldwork periods and outcome definitions. These are human studies, not benchmarks of Morphed or a commercial detector. We did not pool their results or infer a trend across them.
Before you cite these numbers
- September 22, 2026 correction: removed the blanket conclusion that people cannot distinguish current AI images. Added the portrait study as a counterpoint and separated study-specific results from universal claims.
- The prior detection-market and fraud-loss headlines were not supported by sufficiently clear original sources in this review and have been removed.
- Overall accuracy includes both real and synthetic images. The fraction of synthetic images correctly detected and the fraction of participants passing an entire quiz are different measures.
- A preprint is not equivalent to a peer-reviewed result. A 2026 publication date also does not mean that its images came from the latest 2026 models.
- A detector score is evidence about a tool’s classification, not proof of an image’s provenance. Review the original source, available provenance records and surrounding context.
Frequently asked questions
Can people tell if an image is AI-generated?
Sometimes, with accuracy varying by study. Microsoft Research reports 62% in a broad quiz, while a separate portrait preprint reports 85.2%. Neither is a universal rate.
Are AI image detectors reliable?
These human studies do not validate any commercial detector. A detector needs testing on the relevant generators, real-image examples and transformations, with false positives reported.
Why do AI image detection studies disagree?
They use different image pools, model versions, participant samples and tasks. Comparing percentages without those details can be misleading.
Does a 2026 study test the latest AI images?
Not necessarily. The portrait preprint reviewed here used models available in July 2025 and collected responses from August 2025 to January 2026.
Sources
Figures on this page are compiled from the following publishers and reports. Where sources disagree, we present the range and note the methodology difference.