AI Review Algorithm ── The decision to let a machine read promotional material

01Why let a machine read the material

Reviewing promotional material looks like reading, but it is really searching. Is there an expression somewhere that exceeds the approved scope? Has the safety wording been diluted? Is the chart axis honest? The reviewer finds the suspicious spots across dozens of pages, pulls up the exact text of the applicable norm, and puts into words why it is a problem. Most of a reviewer's time goes into this searching and cross-referencing.

The starting point is not efficiency. It is preventing misses. From the first day of design, this system defined the false negative (= a real violation that goes unnoticed) as the gravest failure. A false positive (= suspecting something that is not a violation) is settled by a human check, but a false negative becomes harm on the side of patients and clinicians once the material is out in the world. And we made it possible to measure whether that ranking is being upheld.

02What it does, and what it does not do

What this system does can be summarized in one sentence. It detects, classifies and explains, with grounds traceable to the text of the norm, whether the image formed in the mind of a reasonable reader who has read the material (= the bundle of stated claims and the implications drawn from what is placed next to what and what is left out) stays within the approved facts (= the efficacy, dosage and safety scope approved for that drug, with the electronic package insert as the working reference document). What is compared is not the elements of the material themselves but the image the elements form in the reader. That is the subject of the design philosophy article.

Three things it does not do are also fixed. First, it does not make the final verdict. Whether a correction is needed is decided by the human reviewer. Second, it does not adjudicate medical truth. Whether a clinical trial result is correct is outside the review; only consistency with the approved facts is examined. Third, even when the public eye (= a check that re-reads the material through the eyes of patients, families, journalists, regulators and others) raises a strong warning, the system never stops the material on its own. Warnings go to humans, and the gravest warnings go to management.

03Positioned as first-line screening

The system's judgment vocabulary (= the list of words a verdict may use) includes WHITE (= a verdict of no issue). So if the system returns WHITE, can human review be skipped? No. WHITE is a report that "nothing was detected against the written norms," not a guarantee that "there is no problem." That is why WHITE material is still read by a human.

What is saved is the "time spent searching." Sweeping dozens of pages for suspicious spots, pulling up the applicable norm text, and attaching an explanation of why it is a problem. The machine finishes that much, and the human concentrates on judgment. And that human judgment flows back into a ledger and changes how the next material is searched. First-line screening is the stage before human review, and at the same time the stage after it, learning from it.

04How to read the three articles

Besides this page, the section consists of three articles. We recommend reading them in order.

The first is "Design Philosophy." It explains why we think this way: the axiom that places the object of review on "the image formed in the reader," a second axiom that counts misses, seven principles, and the gates that turn philosophy into enforcement.

The second is "Components." It explains what the system is made of: twelve parts on three planes, the gathering plane, the judging plane and the learning plane, with what each part receives and produces and whether a program or the AI does the work. Figure 2 shows the overall layout.

The third is "Execution Process." It explains what happens in one review, in fifteen steps: from receiving the material to the rule checks, the picture in the reader's head and its differences, the public eye, merging, the human verdict, the panel, and the learning loop back into the precedent notebook and the miss ledger. Figure 3 is the timeline.

Looking at the parts without knowing the philosophy makes no sense, and following the flow without knowing the parts looks like a magic trick. Read in the order philosophy, components, process, and why, what and how connect in a single line.