How it works
What happens after you hit check.
No magic, no black box. Here is what we read in the file, what the model looks at, where it can go wrong, and what happens to your file.
- 1 · In
Your file comes in
Upload a photo or video, or paste a link. It goes to our own servers over an encrypted connection.
- 2 · File
What the file says about itself
Many AI tools sign their files with Content Credentials, AI services in China must write an AI label into what they make, and some apps save their settings in the file. When we find one, we show it as proof, with who signed it.
- 3 · Picture
A model looks at the picture
Two open-source models (MIT license) look at the picture itself: Community Forensics for photos and single frames, and cakelens for how video moves.
- 4 · Verdict
A straight answer
Frame scores roll up into one probability, a plain-English verdict and a timeline you can inspect. When the signal is weak we say inconclusive, and we tell you why.
What it looks for
Cameras capture light. Generators invent it.
A camera records light through a lens onto a sensor. A generator invents every pixel. Both can look real to us, but they don't leave the same traces in the picture.
The models are trained on real photos and footage and on generated ones. For a video, every moment gets a score on the same scale as the verdict, so the timeline shows where it looks most generated.
Reading the score
- Below 26%
- No signs of AI generation. Frame by frame, it reads like camera footage.
- 26–44%
- Inconclusive. Mixed evidence. We won't round it up or down.
- 44–80%
- Likely AI-generated. Clear signs of generation across much of it.
- 80%+
- Very likely AI-generated. Strong, consistent signs across the whole thing.
The 44% and 26% lines are the detector's balanced and conservative operating points.
Accuracy and limits
Good. Not infallible.
We'd rather you trust us a little less and understand us a lot more.
Strongest on original files
It's most reliable on original, high-quality files. Treat every verdict as evidence, not proof.
Re-uploads cost accuracy
Re-uploads and screen recordings strip detail and weaken the signal. Check the highest-quality copy you can find.
Small or dark videos are harder
Tiny or very dark videos give the model less to work with, so a weak signal comes back as inconclusive.
Edits can slip through
Face swaps and lip-sync on real footage can be missed. A low score isn't proof that nothing was changed.
Context still matters
A real photo can still be misleading: old footage, wrong place, cut out of context. Pair the verdict with a source check.
Privacy
Your file is checked, then gone.
- Deleted as soon as the check finishes. Files go to our own servers over an encrypted connection and sit in a temporary file only while the check runs. For links we fetch the public post, check it, then delete our copy.
- Our own servers. We run the detector ourselves, so your file never goes to a third-party AI API.
- No training on your files. What you upload isn't used to train anything.
- You control reports. A report keeps the verdict, the scores, what we found and a few small preview images. Anyone with its link can open it, and its owner can delete it.
The model
Open source, and on our own servers.
Find the AI runs two open-source models (MIT license): Community Forensics for photos and video frames, and cakelens for video. They are strongest on original, high-quality files, and we'll publish our own test results.
Every report records the model version that produced it, so results stay reproducible as the detector improves.
Press or research enquiries: press@findtheai.app