Can Humans Spot AI Images? Real-Image Selection Rates
Across 7,593 recorded exposures of Human-labeled images, the real image was selected 54.1% of the time. This aggregate contains Human-image rows, not complete outcomes for every AI image. Because it also combines game layouts with different choice counts, it is a selection-rate snapshot—not a full real-vs-AI accuracy benchmark.
Results by category
Pick rates vary widely by category. Faces reached 70.2% in this snapshot, while the small singular “Face” category reached 35.3%. These are descriptive values, not fair category rankings: sample sizes, image pools and game layouts differ.
| Category | Human picks | Exposures | Real images | Pick rate |
|---|---|---|---|---|
| Faces | 417 | 594 | 10 | 70.2% |
| Food | 492 | 768 | 20 | 64.1% |
| Animals | 759 | 1,261 | 28 | 60.2% |
| People | 667 | 1,154 | 33 | 57.8% |
| Paintings | 381 | 681 | 24 | 55.9% |
| Examine | 149 | 281 | 4 | 53.0% |
| Buildings & places | 422 | 920 | 17 | 45.9% |
| Sports & action | 131 | 297 | 16 | 44.1% |
| Cars | 306 | 700 | 21 | 43.7% |
| Nature | 345 | 842 | 35 | 41.0% |
| Stuff | 32 | 78 | 35 | 41.0% |
| Face | 6 | 17 | 2 | 35.3% |
Methodology
Snapshot: the public aggregate was computed on August 29, 2026 at 03:15 UTC. It is cumulative; the aggregate does not expose the first contributing event date, so this report does not claim a collection start date.
Unit of analysis: one exposure of a Human-labeled image in a retained first answer. The refresh job keeps the earliest answer per anonymous player, round, media type and content-pool version, then counts each shown image once. Image game modes can have different numbers of choices.
Data source: the public image_stats aggregate generated from image_interactions. For Human-labeled images, exposures counts appearances and picked counts selections. Summing those rows yields 4,107 picks from 7,593 exposures across 245 real images.
Privacy: only aggregate counts are published. The download contains no player identifiers, prompts, IP addresses, or raw event rows.
Limitations
- This is a self-selected website audience, not a representative population sample. “Humans” in the title describes users interacting with the quiz, not a demographic claim.
- Categories contain different image sets and decision counts; category gaps are descriptive, not causal.
- Repeated players can contribute multiple distinct rounds. The refresh keeps only the earliest answer for a given player and round, but it is not a one-response-per-person survey.
- The aggregate combines image game modes with different numbers of choices and exposes no game-mode breakdown. Do not compare 54.1% to one universal chance baseline.
- Local-development writes were not blocked until the production-only persistence guard shipped. Historical development interactions may remain in this cumulative snapshot and cannot be separated in the public aggregate.
- The Wilson interval treats exposures as independent even though observations cluster by player and image; it is descriptive and probably too narrow.
- Generator mix and image difficulty change as the gallery grows. Preserve this dated snapshot rather than treating it as a permanent benchmark.
Try the same task
Play a live image round, or generate an embed code for your own lesson or article.