Welcome to Clairfy

Most AI-detection tools answer a different question than the one you’re asking. You want to know whether you can publish, moderate, or trust the image in front of you. They give you back 0.78. What are you supposed to do with 0.78?

We built Clairfy because explainability is the product. A verdict without reasons is a number you can’t defend. A verdict with reasons is a position you can write down, share, and argue about. That’s what gets a story over the line at a desk, a takedown through an appeals process, or a piece of evidence into a brief.

What a Clairfy verdict actually contains

Every check returns four things:

1. A verdict: AI Generated, Real Photo, or Inconclusive. Three buckets, not a sliding scale — because real decisions don’t move continuously either.
2. A confidence bucket: High, Medium, or Low. We resist showing raw percentages on the front because they imply more precision than the underlying model actually has.
3. A written rationale: a paragraph in plain language explaining why the model called it the way it did.
4. A list of visual signals the model noticed — the specific things in the image that drove the call. These are the receipts.

The rationale and the evidence aren’t decoration. They’re the constraint that keeps the verdict honest. A model that has to commit and explain commits more carefully than a model that only has to commit.

Why a vision LLM and not a dedicated classifier

Classifiers trained on AI-vs-real distinguish reasonably well on the kinds of images they were trained on, and fall over on images they weren’t. The world produces new generators every month. The classifiers don’t keep up.

Frontier vision models trained on a much broader image distribution can be asked to look for the right things, and to flag what they find. That’s a more robust posture against an adversary that’s also improving — and it gives you a verdict you can read.

What this blog will be

Notes from inside the detector:

• What generators give themselves away with (and what they don’t anymore)
• How the prompt evolves and what we learn from misclassifications
• What changes when the underlying vision model gets smarter
• Field notes from journalists, moderators, and researchers using the tool

Not marketing. Not changelogs. Things that would have helped us figure this stuff out a year ago.

Drop in at clairfy.ai/app and try one. Let us know what surprises you.