Skip to content
WatermarkAudit

16 October 2026 · 2 min read

Is there an AI image detector that works?

Pixel-based detectors are near-perfect on the generators they were trained on and much worse on everything else, which is why provenance beats detection on an image you did not make.

It depends what you want it to do. Detectors that judge an image from its pixels are close to perfect on the generators they were trained against and much weaker on anything else, which is the opposite of what you need when a stranger sends you a picture.

What the benchmarks show

The failure mode is consistent across the research. A 2026 benchmark for AI-generated scientific figures reports one detector scoring 97.57% on images from one generator and 28.40% on images from another — worse than guessing, because what it had learned was that first generator's fingerprint rather than anything general about synthetic images. An ICCV 2025 benchmark built from real-world conditions tested 17 detection methods and found none of them saturated, with the best reaching under 90%.

Those conditions matter as much as the model. Compression, cropping, upscaling and screenshotting all strip the fine-grained statistical traces these systems read, and an image that has been through a social platform has had all four done to it.

So a vendor's headline accuracy figure describes a test set, not your file. It tells you how the detector does on generators it has met.

What works better

Reading the file instead of the picture. A C2PA content credential is a signed record of the software that made or edited an image, and a valid one naming an image model is the closest thing to proof of AI origin available today. OpenAI attaches credentials to its images, along with Google's SynthID watermark — though SynthID itself is readable only through Google's own detector portal, which is still waitlisted.

The limit is that provenance is absent far more often than it is present. Platforms strip metadata on upload and a screenshot keeps none of it, so an empty result is the normal state of an image online and says nothing about who or what made it.

If you already have a score

Treat it as a reason to look further, not an answer. Check where the image first appeared and whether anyone else has the same frame. And never use a percentage from one of these tools as evidence against a person — the error rate that matters is the one on the image in front of you, and nobody can tell you what it is.

You can read every provenance signal a file carries with our AI detector.

Check something yourself

Every tool here runs in your browser and is free. See what your files and text actually carry.

All posts