You send an application, a paper, a pitch. Something on the other side decides. You are never told whether that something was a human being.
The usual answer is a detector. Detectors do not work. They are trained on yesterday's output, they flag honest writing as machine writing, and anyone who cares about the result routes around them within a week. Building a better classifier is a race that the person being evaluated always loses.
So the interesting question is not can this text be detected. It is can the reader be made accountable. Two things follow from that, and this site is where they get built in the open.
A short passage you place in a document you are sending. It asks nothing, it changes nothing, and it never tries to influence a decision. It only reveals whether the document was processed by a language model rather than read. If a reply comes back carrying the phrase, you learn something you had no way of learning before.
It is deliberately visible and deliberately harmless. A canary that tries to bias the outcome is fraud, and it would also make the result meaningless.
It changes every day. That is the whole design: a fixed phrase gets memorised, filtered and stripped, so its value decays the moment it becomes popular. A rotating one does not.
Detection asks a machine to judge a machine. Attestation asks a person to put their name on a claim: I read this. I made this. This much of it was mine. That is a social object, not a statistical one. It cannot be defeated by a better model, only by people lying, which is a problem every institution already knows how to price.
What makes it work is the same thing that makes it hard: it is worth nothing until enough people carry it, and everything once they do.
Early, and honest about it. The canary above is real and works today. The attestation layer is a design, not a product, and there is no waitlist to join and nothing to buy. If you want to argue with any of it, that is the most useful thing you could do right now.