How to Review Materials Made by Generative AI── Keeping the pace, while people verify what matters

01The floods of generated materials ── only the making got faster

First, let us put into words what is happening now. By materials we mean the documents used to deliver information about prescription drugs to healthcare professionals and patients ── brochures, explainer slides, web pages, emails, video scripts, and the like, taken together. Traditionally, such materials were finished through repeated back-and-forth: a staff member wrote a draft, put it through internal review, made corrections, and submitted it again. Taking several weeks to complete a single piece was not unusual.

Generative AI shrank this drafting stage to minutes. So what happens? The volume of materials arriving at the review department goes up. Yet the time to verify each piece stays the same. Reading it, checking it against the statutes, tracing the citations, seeing whether it stays within the approved indication ── none of this gets shorter just because AI wrote the draft. Only the entrance got wider, while the exit stays the same width. So it clogs.

Here let us make clear what material review is. Material review is the process of verifying, before publication, that a material does not violate the Pharmaceutical and Medical Device Act (PMD Act), internal standards, or the guidelines for sales information provision activities. And a material is, ultimately, a document for information provision. The PMD Act places the proper conduct of information in sales information provision activities under Article 68-2. Matters of commerce ── price, stock, delivery time, ordering ── are not information to put in a material. That belongs to the domain where pharmaceutical wholesalers and hospital purchasing departments deal with each other, and it is a separate matter from the information provision done by materials or by MRs (= medical representatives). This dividing line will matter later as a fence for review.

02The three traps of mass production ── why "fast" turns into "dangerous"

That materials can be made fast and in bulk is not, in itself, a bad thing. The problem is that generative AI has habits peculiar to how it makes things, and when those multiply with sheer volume, the probability of slipping through review rises. There are three main traps.

What the three traps share is that all of them are "perfect in appearance." A clearly odd material can be stopped by anyone in review. The danger is when a single sentence that casually states an indication beyond the approved range, or a single line attaching a nonexistent citation, sits naturally within a polished material. Claiming an indication beyond the approved range constitutes exaggerated advertising under Article 66 of the PMD Act ── yet that one sentence wears a face natural enough to skim right past.

03The three checkpoints ── a design where people do not read everything

So how do you review a stream of materials that only grows? The answer is not "people re-read everything." That clogs. Nor is it "hand everything to AI." That misses the traps. The realistic answer is to split review into three checkpoints and, after machines narrow things down, have people verify what matters.

• Entrance ── the instruction before it is made: At the stage of the instruction (prompt) that has generative AI make the material, embed the requirements and prohibitions up front. The scope of approved indications and effects, the citations that may be used, the exaggerated expressions to avoid. Handing over the frame here reduces the work of removing things downstream. It is the cheapest checkpoint.
• Middle ── the first sieve right after generation: Pass the material that comes out through mechanical checks. Detection of prohibited words, matching against approved indications, verifying that citations really exist. With AI and rules, drop the obvious deviations here. A coarse sieve before it reaches human eyes.
• Exit ── the final human review: Only what clears the first sieve is read by people. They approve it, record it, and bear responsibility. This checkpoint is not automated. Judgment and signature are, to the end, people's work.

The important point is that in none of the three do people read everything. Fit the frame at the entrance, let machines narrow at the middle, and at the exit have people verify only what matters. To keep the defenses up without slowing down, there is no choice but to narrow down where you verify. People's time is concentrated on what machines cannot pick up ── reading context, subtle assertions, misleading gaps.

04Combining guardrails ── freeing human eyes with machine fences

What supports the middle's first sieve is guardrails. A guardrail is a mechanical fence placed before the output reaches human eyes. In driving terms, it is the curb that physically pushes a car back when it drifts off the road. In material review, line up fences like these.

• Matching against the approved scope ── whether the material's indications and effects exceed the approved range. Exceeding it can constitute exaggeration under Article 66 of the PMD Act.
• Checking the product ── whether it advertises a product not yet approved. Advertising unapproved drugs and the like is prohibited under Article 68.
• Verifying the existence of citations ── whether the cited trial or literature truly exists and matches the content.
• Detecting prohibited words and expressions ── assertions like "the best" or "absolutely safe," and phrasing that stirs expectations beyond what was approved.

There is one more fence: the dividing line of the material itself. A material is a document for information provision, not a tool of commerce. Price, stock, delivery time, ordering, discount negotiation ── these are not information to put in a material. They belong to the domain handled by pharmaceutical wholesalers and hospital purchasing departments, separate from an MR's information provision. Yet when you have generative AI make a document, it sometimes mixes in unrequested promotional lines like "plenty in stock now" or "special price," thinking it is being helpful. Removing such intrusions is also the review fence's job.

The aim of letting machines take on the guardrails is to free human eyes. Detecting prohibited words and cross-checking citations are the machine's forte ── tirelessly and steadily done. Entrust that, and people spend their time on the judgments machines cannot pick up. Fences are placed not to replace people, but to move people to more important places.

05Tailoring for multiple audiences ── change the audience, change the measure

One of generative AI's fortes is tailoring. From a single source, it can make separate versions for physicians, for pharmacists, and for patients. What once took the labor of rewriting for each audience now comes out in parallel from a single instruction. Convenient. But there is a pitfall here. Change the audience, and the regulatory measure you apply changes too.

Patient- and general-public-facing materials call for particular care. Advertising prescription drugs to the general public is, as a rule, not permitted at all. The specialized descriptions allowed in materials for healthcare professionals do not pass in patient-facing ones ── and generative AI does not distinguish this difference automatically. Tailoring mechanically from the same source risks letting physician-facing expressions flow straight into patient-facing ones.

The idea of changing how you provide information according to the audience is not a new one. In its 1988 "Ethical Criteria for Medicinal Drug Promotion," the World Health Organization (WHO) held that promotion should be accurate, fair, and substantiated, and called for consideration according to whether the reader is a prescriber or a member of the general public. Apply that old principle, as is, to generative AI's tailoring too. Being able to make it fast and being able to make it correctly for the audience are separate matters.

06Effect and limits ── AI up to the first sieve, judgment to people

Put this mechanism in place, and review does get faster. Detecting prohibited words, matching citations, cross-checking against the approved scope ── machines handle the first sieve tirelessly, and people need look only at what falls through. You can meet the rising volume without adding more human eyes. This is the effect.

But you need to look honestly at the limits.

• You can only drop what you find ── AI's first sieve picks up and drops the deviation patterns it has learned. Turned around, a new kind of deviation with no precedent can sail through. The "miss" that passes what should be stopped is the most frightening failure.
• "It passed" is not "it is correct" ── clearing the sieve is not proof of the absence of deviation. It is the same structure as Vol. 1's "it worked ≠ it is correct."
• Judgment and responsibility remain with people ── do not let AI stand in for the approval signature. "Because AI checked it" is never an excuse, in any situation.

07Connections to other chapters ── the measure of regulation, and the design of speed

The asymmetry this volume took up ── the "making side" and the "verifying side" ── connects to other chapters on this site. Read together, the place of material review comes into three dimensions.

• AI Material Review series ── the whole picture of generative AI and review. This volume is its practical installment, taking up how to build the three checkpoints and fences.
• Advertising Regulation ── PMD Act 66 / 68 / 68-2 ── the source text of the measure review applies. The basics for not mixing up the three articles.
• Material Review series ── the practice of review that predates AI. Even with generative AI in the mix, the skeleton of the verifying process does not change.
• AI Marketing Vol. 5 ── The Balance of Speed and Review ── the same problem seen from the making side. The way of thinking about a mechanism where people verify the output midway is shared with this volume.

Generative AI changed the speed of making materials by an order of magnitude. But the speed of verifying them has not changed. Leave this asymmetry alone, and one of two things happens: the review department clogs, or the checking goes lax. The answer this volume presented is simple ── stop having people read everything, and narrow down where you verify. Fit the frame at the entrance, let machines narrow at the middle, and have people verify what matters at the exit. Combine the three checkpoints with machine fences, and apply the PMD Act's three measures (exaggeration Article 66, unapproved Article 68, information provision Article 68-2) without mixing them up.

What must not be forgotten is this one point: what AI carries is only up to the first sieve. Judgment and responsibility remain with people to the end. Make the verifying process thicker by exactly as much as you can make faster ── think of the purpose of raising speed as freeing up time to put back into review, and you will not mistake the order. Next time, we step into a particularly troublesome adversary in this review ── how to find, with AI, the deviations that slip into materials. Exaggeration, off-label, missing citations. How far can a machine pick up errors that wear a plausible face?