Detect Deepfakesby Resemble AI

A score is not an answer

Every detector on the market — ours included — returns a probability, not a fact. The number is most useful when you pair it with provenance.

Three questions worth asking before you trust or share a file:

  1. Where did it come from? A file with no traceable origin deserves more scepticism than the score alone suggests.
  2. Who published it first? Reverse-image search and a date check catch a large share of recycled fakes before any model is needed.
  3. Does a primary source exist? If a clip shows a public figure saying something newsworthy, a real recording would exist somewhere else too.

Common questions

How can I tell if an image is AI-generated?
Run it through a detector, then sanity-check the result yourself. Diffusion models still struggle with hands, teeth, reflections, jewellery, and any legible text in the background. Our image detector reports which regions drove the verdict rather than returning a bare percentage.
Is there a free AI detector?
Yes. Every detector on this site is free, requires no signup, and has no per-file limit beyond a 50 MB size cap. The same models are available via the Resemble AI API with a free monthly tier if you need to call them programmatically.
How accurate are AI detectors?
Ours post 93% on images, 94% on video, 98% on audio, and 95%+ on text samples over 250 words, measured on our 2026 benchmark. Accuracy drops on heavily compressed, cropped, or re-encoded media — a screenshot of a screenshot is much harder than an original file.
Can AI detectors be wrong?
Yes, in both directions. Treat a detector as evidence, not proof. Heavily edited real media can read as synthetic, and a high-quality generation passed through compression can read as real. For anything consequential, combine the score with provenance: where did the file come from, who published it first, and does a primary source exist?
Does checking a file store or train on it?
No. Files are processed in memory, deleted within 10 minutes, and never used for training.
What about C2PA and watermarks?
Some generators embed provenance metadata or an invisible watermark. Our watermark inspector reads C2PA manifests and checks for known watermark signatures. Absence of a watermark proves nothing — most models do not add one, and metadata is trivially stripped — but its presence is strong positive evidence.