How accurate is a free AI image detector?
The short answer
A free detector is as accurate as its model and its method, not its price. In our test the detector alone flagged 85% of AI pictures and 25% of real ones. Combined with the file, the size and the noise, pictures scoring 80 or more were AI 89% of the time and those under 35 still 14%. Those shares come from a test where 57% of pictures were AI, so on mostly real photos a high score is wrong more often. Read a detector's percentage as a hint, with its error rate next to it.
"96% accurate" on a detector's page means little until you know what it was tested on and what the mix of pictures was. This guide gives our numbers for the free AI image detector on this site, and shows how to turn any detector's output into something you can use.
What we measured
We tested on 249 pictures (142 AI from 18 services, 107 real), each as the original file and as a reposted copy (shrunk to 1080 pixels, JPEG quality 80, metadata gone), 498 test pictures in all. To avoid flattering the result, the combined score was trained with whole generators and whole kinds of real picture held out, and the numbers below come from the held-out pictures only.
| Score | Pictures | Really AI |
|---|---|---|
| Under 35 | 184 | 26 (14%) |
| 35 to 65 | 66 | 44 (67%) |
| 65 to 80 | 83 | 67 (81%) |
| 80 and over | 165 | 147 (89%) |
Why the mix of pictures matters
57% of our test pictures were AI. If you check a pile where 5% are AI, most pictures that score high will be real ones the detector got wrong: a detector that wrongly flags a quarter of real pictures flags about 24 of 95 real pictures for every 4 AI ones it catches out of 5. The same score means less when AI is rare. Our percentages describe our test, not your pile, so use them as a ranking, not a probability.
Originals versus reposted copies
Held-out AUC was 0.90 for originals and 0.85 for reposted copies. The file evidence is gone from a copy and the size no longer matches a generator's, so expect less from anything that has been through social media. The detector alone barely changed: it flagged 85% of AI pictures in both.
How to read a detector's percentage
- Find the error on real pictures. A detector's headline is usually the catch rate. The false-flag rate decides whether it is safe to act on.
- Think of the mix. If AI is rare where you work, trust a high score less.
- Prefer file evidence. A prompt, Content Credentials or a named generator needs no percentage.
- Never use a score to accuse. Report what you saw in the file or found on a search.
What free does not change
Some marks (Google's SynthID, OpenAI's and Meta's) can only be read by their makers, a stripped file says nothing, and drawings fool detectors: ours called all five real anime pictures AI, so it gives no number for flat-color art. See our comparison of detector models for where each fails.
A free AI image detector with no sign-up: check as many pictures as you like, with the reasons and the real accuracy numbers. It runs in your browser. Open the Free AI Image Detector.
Frequently asked questions
Is 89% the chance that a high-scoring picture is AI?
Only in a pile like our test, where 57% were AI. In a pile with fewer AI pictures the real chance is lower. Treat the figure as a ranking.
Why is the lowest band still 14% AI?
Stripped files, photo-real services such as Firefly and edited pictures can look unremarkable to a detector. A low score never shows a picture is real.
Are reposted pictures harder?
Yes. The file evidence and the generator size are gone, and held-out AUC fell from 0.90 to 0.85 in our test.