AI Image Detector Accuracy: Our 100-Image Test of Four Detectors

On October 13, 2026 we ran 100 images through four commercial AI image detectors: 50 real photos and 50 AI images from six generator families. Sightengine was the most accurate at 99%, catching 49 of 50 AI images without flagging any of the 50 real photos. Hive came next at 95%, also with no false alarms, but it missed more FLUX images. AI or Not and WasItAI ran only on small samples, and both flagged real photos as AI: 1 of 10 and 2 of 5.

This page reports AI image detector accuracy from that test: the sample, the method, the results by generator and the limitations. The flow our AI image detection checker uses, Content Credentials first and then Sightengine, judged 99 of the 100 images correctly. Our checker now uses Sightengine as its main engine and Hive as its backup; the disclosure at the end explains what that means for how you read this report.

The test set has 100 images: 50 freely licensed real photos and 50 AI images from six generator families, 42 of them from newer generators. It mixes original files with the kinds of copies people actually check, such as screenshots, recompressed images and lightly retouched photos, and every image has a recorded source and license.

Real photos (50). Freely licensed photos, mostly from Wikimedia Commons, under the licenses listed above. 38 are originals. The rest are copies we made to mimic how photos travel: 4 phone-style screenshots, 4 heavily recompressed copies of the kind chat apps produce, and 4 lightly retouched versions. 37 of the 50 kept a camera model in their EXIF, from iPhone, Pixel and Galaxy phones.

AI images (50). Six generator families, with every model listed in the results below. 42 come from newer generators, meaning everything except Stable Diffusion XL and Midjourney v6. Some we generated for this test; others come from openly licensed public datasets. 10 are phone-style screenshots of AI images, saved as PNG with no metadata at all.

Each engine received the same files, and we turned its answer into AI or real with one fixed rule per engine: a Sightengine AI score of 0.5 or higher, Hive's recommended threshold of 0.9, and the vendor's own verdict for AI or Not and WasItAI.

Accuracy is the share of tested images judged correctly; an "unknown" answer counts as wrong. The number we watch most is real photos flagged as AI, because every false alarm lands on a real person: a buyer, an artist, a date. Even with 0 false alarms in 50 real photos, the 95% confidence interval runs from 0% to 7.1%.

Content Credentials were read separately, with the Content Authenticity Initiative's open-source C2PA library. We also scored the flow our checker actually uses: credentials first, then Sightengine for every image they didn't settle.

FLUX was the hardest family for both main engines: Sightengine caught 8 of 9 FLUX images and Hive only 6 of 9. Midjourney was caught every time by both engines but never carried Content Credentials, so it always needs a model check.

The credentials column reflects the files we had. Some came from public datasets, where a re-saved file can lose its credentials, so a lower count doesn't mean that generator never signs its images.

On their own, Content Credentials identified 24 of the 50 AI images, all of them original files, signed by OpenAI, Google, Black Forest Labs or BytePlus (ByteDance). None of the 10 screenshots carried credentials. Credentials never pointed at a real photo; the only real photos with credentials were camera credentials from a Google Pixel 9 and a Google Pixel 10.

Screenshots cost the engines a little. Across 10 pairs of an original and its screenshot, Sightengine went from 10 detected to 9, and Hive from 10 to 8.

This is a small test of AI image detector accuracy, run over one day by a site that now uses two of the engines, and it has gaps. Read every number on this page with these limitations in mind.

• It is small. 100 images is enough to spot a bad engine, not to pin down a good one. Zero false alarms in 50 real photos can't rule out a rate of a few percent.
• Uneven samples. AI or Not ran on a stratified sample of 20 images and WasItAI on 9, because we stayed within their free allowances. Treat those two rows as early signals, not rankings.
• One late image. One real photo went through Sightengine and Hive a few hours after the rest, once their daily free quotas had reset. Both called it real.
• What's missing. No dating-app selfies or heavily filtered portraits, no hand-made artwork among the real images, only a few illustrations among the AI ones, and no composites or partial AI edits.
• A snapshot. Vendors update their models, so these results can change. List prices are the vendors' own figures at the time of testing; every run used free trial credits.

Based on these results, our checker sends images that need a model check to Sightengine, and to Hive if Sightengine fails or times out. Content Credentials are read in your browser first, and when they already say AI, no engine is called at all.

That gives us a stake in these two engines looking good, which is why the full per-engine numbers, our decision rules and the limits are all on this page. We ran all four engines on free trial credits; no detection company paid for this test or had a say in it. We didn't pick AI or Not partly because its pricing page reserves reseller rights for its Business plan, and we didn't pick WasItAI because it flagged 2 of 5 real photos in our small sample.

Sightengine, at 99% across all 100 images, with no real photos flagged as AI. Hive was next at 95%, also without false alarms. The other two engines were tested on samples too small to rank them properly.

Sightengine and Hive: 0 of 50 each. AI or Not: 1 of 10. WasItAI: 2 of 5. Small samples swing a lot, but a false alarm on a real photo is the error we care about most.

No. A test tells you how often an engine was right on images like ours. Your picture may be a screenshot, an illustration or a partial edit, which we tested little or not at all. That is why every answer comes as a range with reasons.

We ran it ourselves and we now use two of the engines we tested, so it is not independent in that sense. No vendor paid for it or saw it in advance, and we publish the full numbers, the method and the limits so you can judge for yourself.