The best AI image detectors, and what each one gives you

Updated 2026-09-08 · 7 min read

No single winner comes out of this, because the tools people compare here do two unrelated jobs. One kind reads what the file records about itself and hands you evidence. The other looks at the pixels and hands you a percentage. Run the first kind first. The second is a second opinion, and on some images it is barely that.

Two different tools, one name

A provenance reader looks for a signed C2PA manifest, an IPTC source type, an XMP creator tool field, EXIF capture data or a watermark. What it finds was written by the software that made or edited the image, so a positive result is a record rather than an inference. The limit is severe: platforms strip metadata on upload, so most images arriving from social media have nothing left to read.

A classifier ignores all of that and scores the picture. It works on any file, including a stripped one, which is exactly why people reach for it. What it gives back is a probability produced by a model whose training data you cannot see, on an image whose compression history you also cannot see.

Provenance readers, all free

  • The image check on this site. Reads C2PA, IPTC, XMP, EXIF and PNG parameter blocks in the browser; the file is never uploaded anywhere.
  • Content Credentials Verify, run by the Content Authenticity Initiative. Opens the signed manifest in full: who signed it, which tool, which edit steps, in order.
  • SynthID Detector from Google DeepMind, for the watermark Google embeds in Imagen, Gemini, Lyria and Veo output. The portal is gated, and the Gemini app answers the same question for an image you upload to it.

Pixel classifiers, and what each returns

ToolWhat you get
HiveA public demo plus a moderation API. It returns a confidence between 0 and 1 and names the engine it believes produced the image, from a list that includes Midjourney and Stable Diffusion XL. The named engine is the useful part: it can be checked against what the picture claims to be.
SightengineA browser demo and a developer API with the same model attribution idea, images only. Built to be called from code, so the free demo is a shop window rather than the product.
DeepAIFree, no account needed for the basic check, one number back. The thinnest explanation of the group, and the easiest thing to run when you have thirty seconds.
WasItAIOne page, one upload, one verdict, no signup. It ranks near the top for most of these searches and it is fast, but a verdict without a reason is difficult to act on.
Winston AIA paid product with a trial and a credit system. Winston says its image report combines the classifier score with a read of C2PA and IPTC metadata and full EXIF, which is the one commercial report here that contains something you can go and verify yourself.

Why no accuracy numbers appear on this page

Every vendor above publishes a figure. Those figures come from the vendor's own test set, and a detector evaluated on the data it was tuned for is a detector at its best.

There is a public measurement of what happens away from that. Deepfake-Eval-2024 collected media that actually circulated in 2024: 45 hours of video and 1,975 images from 88 websites in 52 languages, drawn from social platforms and from users of detection services. Against that set, the area under the curve for open-source detectors fell by about 45% on images and about 50% on video compared with the academic benchmarks the same models score well on.

We have not run our own bench on these products, so we publish no number for any of them. The measured field in our detector database stays empty until that changes. Our methodology page says what the bench will have to do before a figure goes anywhere near this site.

Choosing between them

QuestionAnswer
If I only use one thing?Read the file. It costs one drag and drop and it is the only route to an answer you can show somebody else.
The file is stripped. Now what?Run two classifiers rather than one. Agreement is weak support. Disagreement means you have no answer, which is worth knowing before you act on one.
Is there a genuinely free AI image detector?Yes for provenance, everywhere. For classifiers, DeepAI and WasItAI take an image without an account, and Hive and Sightengine both run public demos.
Do they work on a screenshot?Worse than on the original. A screenshot re-encodes the picture and discards the metadata, and re-compression is the condition these models degrade under.
Can I use a result as proof?A classifier output, no: it is an opinion with an error rate. A signed C2PA manifest, yes: it is checkable by the person you are arguing with.

The short version

Rank these tools by what they hand back, not by the accuracy each one advertises. Provenance readers are free and give evidence when there is any left in the file. Classifiers cover the rest and return a probability whose real world performance, on the one public in-the-wild benchmark, drops sharply from the published figures. Hive and Winston at least tell you why they answered as they did.

Sources

Read next