What we check

Four kinds of signal, and an honest column for each one.

This is the page we would want to read if we were buying this. What we can actually do when you send us a document today, what we cannot, and the three things people keep expecting us to be.

What we check, and what actually works today

Where each one stands today

Read the third column first. It is the one most sites in this market do not print.

SignalWhat it would tell youWhere we stand today
Content Credentials (C2PA)A signed record attached to the file saying how it was made and by what.Working. We read these on PDFs and on text, check the signature against the official trust list, and reject a file where one character has changed.
Google SynthIDA hidden marker some AI systems place in text they generate.Partly. We can detect it where the keys are known to us, which is how we test and calibrate. We cannot read Google’s production watermark, because that needs Google’s own detector.
Claude’s text watermarkThe same idea, for text written by Anthropic’s models.Not yet. Detection is in private preview and we have not been granted access, so we cannot detect it today. The code is written and waiting. We will say so here the day that changes.
Statistical detectorsA guess, from the writing style, about whether AI was involved.Supported, deliberately limited. You can connect one you already pay for. We measure how often it is wrong and weight it accordingly. It can never produce an accusation on its own.

If that looks like a short list, it is. A watermark only exists if whoever made the content chose to put one there, and most of the world has not started yet. The useful product is the one that is straight with you about that.

What we are not

Three things people expect us to be, and we are not.

NOT THIS

A plagiarism checker

Plagiarism tools ask whether your writing resembles something that already exists. That needs a library of other people’s documents. We do not hold one and we do not compare your work to anyone else’s.

NOT THIS

A similarity report

We do not produce a percentage of matched text. If that is what you need, you need a plagiarism tool, and we will tell you so.

NOT THIS

An AI-writing detector

Those score how AI-sounding your prose is. Independent testing puts them well below their advertised accuracy, and they misfire hardest on people writing in a second language. We read evidence the document actually carries. Where we do use a detector, we cap how much it counts.

Why we publish this

Because the alternative is that you find out later, and because a verification company that oversells its own coverage has destroyed the only thing it was selling. If a check starts working, this page changes and the date changes with it. If one stops working, the same.

How these checks are combined into an answer →