A text produced by artificial intelligence can be convincing even when it contains an inaccurate reference. Verification should focus on the elements that the reader could interpret as facts: names, dates, figures, sources and descriptions of obligations.
I propose to separate style review and content control. A link should be opened and compared with the sentence it should support. The existence of the source does not prove that that specific statement is contained in it. Quotations and titles may also require direct verification.
A hypothetical example is a synthesis that transforms a prediction into an outcome that has already taken place. Control must recover period and degree of uncertainty, in addition to the number. In sensitive texts, the reviewer’s competence and access to sources are important conditions.
The NIST AI Risk Management Framework proposes a voluntary reference for AI risk management. Here the editorial procedure is an operational proposal: classify statements, verify the material ones and correct the text before use. If information cannot be confirmed, it must be amended or the limit made clear. The desired result is a verifiable content, the quality of which does not depend solely on the fluidity of the writing.