AI Image Detection Statistics: Can People Tell What’s Real?
AI image detection: human accuracy ranges 49–62% across major studies (chance = 50%); only 0.1% of iProov participants reliably distinguished all real vs AI content. Deepfake detection market: $5.5B (2023) → $15.7B (2026), ~42% annual growth.
The scientific consensus is now firm: humans cannot reliably distinguish AI-generated images from real photographs. Across studies covering hundreds of thousands of evaluations, accuracy clusters between 49% and 62% - at or barely above coin-flip. That failure is fueling a deepfake detection market growing at roughly 42% per year.
Key statistics
Analysis of ~287,000 image evaluations by 12,500+ participants found 62% overall accuracy - best on portraits, significantly worse on landscapes and urban scenes.
In a 1,276-participant study, mean detection accuracy for synthetic images was 49.4% - below random chance. Overall synthetic media detection averaged 51.2%.
iProov research found only 0.1% of participants could reliably distinguish real from AI-generated content across all formats.
The deepfake detection market is growing from $5.5B (2023) to $15.7B (2026) - roughly 42% annual growth.
Deepfakes account for ~11% of global fraudulent activity in 2026 - a category that didn’t register before 2022.
Every major human-detection study, side by side
Five independent studies in 2025–2026 measured human ability to spot AI images. None found reliable detection; the best result (62%) barely beats a coin flip, and the worst (29% on FLUX images) means people actively believed AI images were real.
| Study | Sample | Accuracy | Key finding |
|---|---|---|---|
| Microsoft Research | ~287K evaluations, 12,500+ people | 62% | Worst on landscapes; best on portraits |
| Communications of the ACM | 1,276 participants | 49.4% (images) | Below chance for synthetic images |
| arXiv portrait study (2512.22236) | 165 users, 233 sessions | 54% | No improvement with repeated attempts |
| Frontiers in AI (FLUX.1-dev) | Mixed-methods 2025 | 29% on FLUX images | Below chance: AI judged more "real" than photos |
| iProov | Multi-format testing | 0.1% reliably correct on all | Practically nobody passes consistently |
All five studies linked from the stats above; accuracy figures are each study’s headline result, not directly comparable due to differing image sets.
What do the detection studies actually show?
Every major 2025 study converges on the same conclusion - and newer models make the problem worse, with FLUX.1-dev images detected at below-chance rates.
165 users across 233 sessions classified curated Midjourney portraits vs real photos at just 54% accuracy, with no improvement across repeated attempts.
Participants correctly flagged FLUX.1-dev images as AI in only 29% of cases - below chance, meaning they actively believed AI images were real.
Human detection rates for high-quality video deepfakes sit at just 24.5%.
Gartner predicts 30% of enterprises will no longer consider standalone identity verification reliable in isolation by 2026.
How we compiled this data
This page cites each detection study directly (Microsoft Research, ACM, arXiv, Frontiers, iProov) rather than secondary roundups, and reports each study’s own headline accuracy with its sample size. We present them as a range deliberately: image sets, participant pools, and exposure times differ, so a single blended "humans detect at X%" number would misrepresent the literature. Last full review: June 12, 2026.
Before you cite these numbers
- Study results are not directly comparable: each used different models, image curation, and time limits. Cite a specific study, not an average.
- Detection accuracy decays with model generation; 2025 results on FLUX/Midjourney v6 already understate how undetectable 2026 models are.
- The deepfake detection market figure ($15.7B) includes identity-verification infrastructure, not just image-classifier tools.
- Lab studies measure unprimed observers; accuracy improves when people are warned and given time, which most real-world feeds do not provide.
Frequently asked questions
Can humans detect AI-generated images?
Not reliably. Study accuracy ranges from 49.4% (below chance) to 62% across major 2025 research. Only 0.1% of participants in iProov testing could reliably distinguish all real from AI content.
Which AI images are hardest to detect?
Natural and urban landscapes fool people most (Microsoft study), and images from newer models like FLUX.1-dev were detected at below-chance rates (29%). Human portraits are slightly easier to judge.
How big is the deepfake detection market?
Approximately $15.7 billion in 2026, up from $5.5 billion in 2023 - roughly 42% annual growth, driven by deepfakes reaching 11% of global fraud activity.
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.