Photo Intelligence Lab by NarkisBlind testMethodology
AI or real photo blind test · live results
Can people tell AI photos from real photos?
Be one of the first players
The live score appears after the first 30 answers.
Photos called right
91.7%
95% interval 64.6%–98.5%. A coin flip scores 50%.
Players
1
1 finished every photo in their round.
Answers counted
12
0 answers faster than 300 ms were left out.
Live, preliminary data: a dated snapshot is published at 1,000 counted answers (12 so far). Updated .
Scores
How players scored
| Accuracy | Rounds |
|---|---|
| 0–9% | 0 |
| 10–19% | 0 |
| 20–29% | 0 |
| 30–39% | 0 |
| 40–49% | 0 |
| 50–59% | 0 |
| 60–69% | 0 |
| 70–79% | 0 |
| 80–89% | 0 |
| 90–100% | 1 |
Photos
How often each photo fooled players
Each bar is one photo in the set, most convincing first. We don't say which photos are AI here, and don't show them: that would give the answers away to future players.
- Photo 1100% n=1
- Photo 20% n=1
- Photo 30% n=1
- Photo 40% n=1
- Photo 50% n=1
- Photo 60% n=1
- Photo 70% n=1
- Photo 80% n=1
- Photo 90% n=1
- Photo 100% n=1
- Photo 110% n=1
- Photo 120% n=1
Methodology
How the numbers are made
- Photos
- Everyday photos of real people, mostly self-portraits they shared with consent or published under a free licence, and AI photos of other people, so no one is shown both ways. Every photo gets the same crop, size and re-encoding with all metadata removed, and is served under an opaque URL. The method page lists every source.
- Rounds
- Each round deals 12 photos at random from the set, 6 AI and 6 real, in random order. The answer for a photo is revealed only after it is answered.
- Sample
- Anyone can play; players are self-selected visitors, not a representative panel. Age band, gender and country are optional and self-reported. A player is a round with at least one counted answer.
- Exclusions
- Answers given less than 300 ms after the photo was served (timed by our server) are treated as likely bots and left out. Developer test rounds are left out. A player can answer each photo in a round once; repeat answers don't change the first one. The score distribution only uses finished rounds.
- Uncertainty
- The 95% interval is a Wilson score interval over all counted answers. It treats answers as independent, but answers are grouped by player, so the true uncertainty is somewhat wider. The dated snapshot will use a player-clustered bootstrap interval.
- Privacy and interest
- No IP address is stored, only a hash of the browser's user agent and the country reported by our CDN. Narkis sells AI photos and so has an interest in the result; the method and the uncertainty are published for that reason.
The aggregated results are published under CC BY 4.0 by the Photo Intelligence Lab by Narkis. Snapshot: published when the study passes 1,000 counted answers. Full method: where every photo comes from, how it is prepared, and what we won't claim.
Embed the live test on your site
<iframe src="https://narkis.ai/research/blind-test/embed" title="Can you tell AI from a real photo? Blind test by the Photo Intelligence Lab by Narkis" width="100%" height="260" style="border:0;max-width:560px" loading="lazy"></iframe> <p><a href="https://narkis.ai/research/blind-test">AI or real photo? Blind test</a> by the Photo Intelligence Lab by Narkis</p>