PROV-03 · Disclosure that readers notice

Write one AI-assistance disclosure three different ways, put them in front of real readers, and measure what those readers actually understood. Compliance and comprehension are different results — which one does your current disclosure achieve?

One real or realistic piece from your publication with a genuine AI-assisted element to disclose. Three disclosure treatments varying one thing at a time — placement, wording, or prominence — designed before any reader sees them. At least twelve readers from your actual audience, none of whom work on the publication, each seeing one treatment. A short comprehension check written in advance: what was AI-assisted, what was not, and would they trust the piece.

Asking readers whether they noticed the disclosure, which tells you about the question rather than the design. Varying three things at once so no result is attributable. Using classmates who already know the answer. Treating "I would trust it less" as failure — an informed reader trusting a piece less can be the disclosure working. Declaring a winner from twelve readers as though it were a finding about all readers.

A disclosure test: the three treatments as readers saw them; the comprehension results per treatment with counts, not percentages of twelve; a written account of what readers misunderstood, quoted; a recommendation with its sample limits stated plainly; and the treatment your publication will adopt, with the reason.

CT: distinguishing compliance from comprehension. CR: designing a fair single-variable comparison. CO: readers are the instrument; their misreadings are the data. CM: the disclosure itself. CZ: the reader who acts on a piece they misunderstood.

A standards editor, designer, or audience-engagement journalist reviews the treatments and the comprehension check before testing. Standing instruction: reject a design that varies more than one thing at a time. School-supervised.

R1: one piece, one variable, three treatments. R2: procedure and raw responses recorded. R3: comparator is the publication's current disclosure. R4: sample limits stated; misreadings quoted. R5: the adopted treatment is ready to use. R6: refuses to generalize from twelve readers.

(a) "Did you see the label?" — a question that measures the question. (b) Percentages from twelve readers presented as audience research.

Lane A immediately (journalism, or a design or statistics cross-listing). No verified credit claim.