# Tested four AI image detectors on 30 images. Here is what the numbers showed **URL:** \ **Category:** Experiments & Case Studies\ **Created:** [17 September 2026 00:53 UTC](https://humanizeaiforum.com/t/tested-four-ai-image-detectors-on-30-images-here-is-what-the-numbers-showed/215 "2026-09-17T00:53:04Z")\ **Posts on this page:** 6\ **Page:** 1
**Author:** ![Appartement](https://yyz1.discourse-cdn.com/flex033/user_avatar/humanizeaiforum.com/appartement/32/34_2.png) [@Appartement](https://humanizeaiforum.com/u/Appartement)\ **Post date:** [17 September 2026 00:53 UTC](https://humanizeaiforum.com/t/tested-four-ai-image-detectors-on-30-images-here-is-what-the-numbers-showed/215/1 "2026-09-17T00:53:04Z")
tested four AI image detectors on 30 images. here is what the numbers showed. 20 AI-generated images across different models and styles. 10 human-created images in styles that overlap with common AI aesthetics (digital art, photorealistic illustration). same 30 images through each tool, same settings. ---
**Author:** ![penhillcurrant](https://yyz1.discourse-cdn.com/flex033/user_avatar/humanizeaiforum.com/penhillcurrant/32/23_2.png) [@penhillcurrant](https://humanizeaiforum.com/u/penhillcurrant)\ **Post date:** [18 September 2026 12:23 UTC](https://humanizeaiforum.com/t/tested-four-ai-image-detectors-on-30-images-here-is-what-the-numbers-showed/215/2 "2026-09-18T12:23:07Z")
the stylistic overlap category is where I’d expect the most noise. human-created work in AI-adjacent aesthetics is the hardest case and probably the most important one for practical use. ---
**Author:** ![strawtin](https://yyz1.discourse-cdn.com/flex033/user_avatar/humanizeaiforum.com/strawtin/32/26_2.png) [@strawtin](https://humanizeaiforum.com/u/strawtin)\ **Post date:** [18 September 2026 19:58 UTC](https://humanizeaiforum.com/t/tested-four-ai-image-detectors-on-30-images-here-is-what-the-numbers-showed/215/3 "2026-09-18T19:58:07Z")
running the same set through four tools and comparing is more useful than any single-tool accuracy claim. the disagreement pattern tells you as much as the individual scores. ---
**Author:** ![maya.glitch](https://yyz1.discourse-cdn.com/flex033/user_avatar/humanizeaiforum.com/maya.glitch/32/39_2.png) [@maya.glitch](https://humanizeaiforum.com/u/maya.glitch)\ **Post date:** [19 September 2026 20:40 UTC](https://humanizeaiforum.com/t/tested-four-ai-image-detectors-on-30-images-here-is-what-the-numbers-showed/215/4 "2026-09-19T20:40:08Z")
the results on the human-created images specifically: that false positive rate is the number that matters most for anyone using these tools to make decisions. ---
**Author:** ![tallest.tower.2016](https://avatars.discourse-cdn.com/v4/letter/t/c89c15/32.png) [@tallest.tower.2016](https://humanizeaiforum.com/u/tallest.tower.2016)\ **Post date:** [21 September 2026 05:17 UTC](https://humanizeaiforum.com/t/tested-four-ai-image-detectors-on-30-images-here-is-what-the-numbers-showed/215/5 "2026-09-21T05:17:30Z")
overall accuracy (correct classification out of 30): Tool A: 23/30 (77%) Tool B: 21/30 (70%) Tool C: 19/30 (63%) Tool D: 18/30 (60%) false positives on human-created images (flagged as AI): Tool A: 2/10 (20%) Tool B: 3/10 (30%) Tool C: 4/10 (40%) Tool D: 4/10 (40%) tool agreement on all 30 images: 17/30 (57%): the four tools agreed on just over half. the false positive rate on human-created work in AI-adjacent styles (digital art, photorealistic illustration) was the key finding. the best tool flagged 2 in 10 human images as AI. the worst flagged 4 in 10. any policy or decision that treats a single tool’s output as determinative is operating on a false positive rate of 20-40% on the most relevant image category. that’s not usable for consequential decisions. ---
**Author:** ![ErnestPrice](https://yyz1.discourse-cdn.com/flex033/user_avatar/humanizeaiforum.com/ernestprice/32/24_2.png) [@ErnestPrice](https://humanizeaiforum.com/u/ErnestPrice)\ **Post date:** [22 September 2026 08:48 UTC](https://humanizeaiforum.com/t/tested-four-ai-image-detectors-on-30-images-here-is-what-the-numbers-showed/215/6 "2026-09-22T08:48:29Z")
the 57% agreement rate across four tools is the number that should give anyone pause. the tools agree on barely more than half the images. on the ones they disagree about, what’s the right answer? we don’t have a ground truth method that would tell us.