Learn
Guides to AI detection
How AI content gets labelled, what detectors can and cannot do, and how to check things yourself when the stakes are high.
All guides
Three ideas that make everything else easier
Provenance beats detection
When a file carries a signed record of how it was made, you don’t need to guess. Content Credentials, IPTC labels and embedded generator settings turn “probably AI” into “declared AI”. The catch is that this information is easily lost in screenshots and uploads, so its absence means very little.
Detection is probability, not proof
Every content-based detector — for images, text, audio or video — produces an estimate that is wrong some of the time, and wrong more often on content unlike its training data. Good detectors say how confident they are and why. Treat a single score as a reason to look closer, never as a verdict about a person.
Context is the strongest check of all
Most misleading media is not generated at all: real photos with false captions, real clips from another year, real quotes stripped of context. Finding the original source, the date and corroborating material answers those cases and helps with AI ones too. Our “Is this photo real?” page puts the steps in order.
How we measure our own tools
Each detector page includes the accuracy we measured on labelled data, broken down by source, and the methodology page explains the data and the method. We tune our tools to keep false accusations of human work rare, which means they miss some AI content. We would rather say “inconclusive” than accuse a student, a photographer or an artist on weak evidence.
What our detectors can’t tell you
No detector — ours or anyone else’s — can tell you whether a claim is true, whether a quote was said, or whether a real photo shows what its caption says. They answer a narrower question: what does the content itself reveal about how it was made? The guides here help with the wider question too, by explaining where to look for sources and context.
They also explain the limits plainly. Watermarks only cover the companies that add them; metadata disappears in screenshots; classifiers can be fooled by new generators and can misjudge unusual human work. Knowing those limits is what lets you use the tools well: as a quick first check, a way to find what to look at more closely, and a source of concrete evidence when a file does declare its origin.
If you are new to this, a good order is: how to tell if an image is AI-generated, what C2PA is, then the SynthID guide. Teachers will find the essay checker page useful, and journalists the fake news photo checker.
Each guide links to the detector it relates to, so you can move from reading about a check to running it on your own file in one click.