Provenance is a reader problem, not a label problem

The first note was about what a summary drops. This one is about the other half: what a reader does with what is left, and why a label alone rarely decides it.

Most work on AI provenance is about the mark: a watermark, a metadata field, a badge above the text. The engineering question is whether the mark travels intact. The human question is what happens when someone reads it, and that question gets far less attention than it deserves.

A label is not an instruction. Told that a passage is machine-generated, one reader discounts it entirely, another assumes the machine had more information than the person, and a third does not register the label at all because they are reading to answer a clinical question at four in the afternoon. Same mark, three outcomes, none of them encoded in the badge.

Where the effect probably lives

My suspicion is that the label matters less than the reader's prior about the task. If someone believes that generated text is a draft, they edit it. If they believe it is an answer, they forward it. The label changes the prior only in the presence of time and attention, and professional reading is exactly where time and attention are scarce.

That suggests a different way to evaluate provenance: not "does the mark survive the pipeline," but "given this mark, what does a reader do next, and is that action the right one." Those are different measurements, and the second one requires readers rather than logs.

A provenance mark is a message. Treating it as a property of the artifact is why the messages fail.

It also means provenance is a communication problem before it is an infrastructure problem, which is the reason I keep ending up on this side of it.