AI and Material Review ── What Is Actually Happening── What changes, and what does not. Drawing the map of all 10 parts before we begin

01The Work of Material Review ── The Last Checkpoint Before It Leaves the Company

First, let us pin down this work in words. Material review is the work of internally checking, before it leaves the company, the materials a pharmaceutical company hands to healthcare professionals ── product information brochures, slides, web pages, emails, and the explanatory printouts used by MRs (= medical representatives). A pair of eyes other than the creator's confirms each one, sheet by sheet. What is confirmed boils down to roughly the following five points.

• Does it stay within the scope of approval? ── Do the indications and the dosage/administration go beyond what was approved?
• Is it exaggerated? ── Are there expressions that overstate efficacy, convenient emphasis of favorable numbers, or phrasing that implicitly suggests superiority over other drugs?
• Does it suggest an unapproved indication? ── Does it quietly hint at a use that has not yet been approved?
• Is safety information presented alongside? ── Are adverse reactions and precautions presented in a manner balanced against the efficacy claims?
• Are the citations accurate? ── Do the cited data and references actually exist, and is the way they are described consistent with the original source?

The fence around this checking is the Pharmaceuticals and Medical Devices Act (= the Act on Securing Quality, Efficacy and Safety of Products Including Pharmaceuticals and Medical Devices; hereafter the PMD Act). The prohibition of exaggerated advertising sits in Article 66, the prohibition of advertising unapproved pharmaceuticals and the like in Article 68, and the provisions on proper information provision in sales information provision activities and the like in Article 68-2. On top of these, the commentary on the Standards for Fair Advertising of Drugs and the Like issued by the Director of the Compliance and Narcotics Division, Pharmaceutical Safety and Environmental Health Bureau, Ministry of Health, Labour and Welfare (MHLW), and the 2018 Guidelines for Sales Information Provision Activities (hereafter the SIP Guidelines) serve as the measuring sticks for judgment.

02Where AI Stands Now ── Not "AI That Reviews" but "AI That Helps Review"

So what is AI actually doing on the floor now? To be precise, AI is not taking over the review itself. What has entered are assistive tools that speed up the reviewer's hands. Writing drafts, searching past findings, marking suspicious expressions ── these three are the uses actually in operation.

Let us be honest about one thing here. For AI built into medical devices and AI used in drug discovery, bodies such as the FDA (= the U.S. Food and Drug Administration) have set out their regulatory thinking in documents. But dedicated regulatory guidance for using AI in material review itself is not yet in place. So the floor decides its own operational rules by drawing on neighboring domains ── the thinking on managing medical-device AI, and the general limits of generative AI. Let us share this current position ── that the guidance is still thin ── up front.

03Three Things AI Can Do ── Create, Detect, Verbalize

The work AI genuinely helps with around material review sorts into three broad kinds: create, detect, verbalize. To these we add "tidying up" as preparatory work, making four. The core remains the first three.

What the three share is that each stops at a "draft" or a "candidate." The text created is a draft, the phrases picked up are a list of suspicions, the reasons verbalized are a draft explanation. Whether to adopt or discard them in the end returns, as described next, to human judgment.

04What Changes and What Does Not ── Speed Changes, Responsibility Does Not

When AI enters, what about material review changes, and what does not? Because this is easy to confuse, we split it out in a table. The left is where AI genuinely changes things; the right is what remains with humans no matter how clever AI becomes.

The point is a single one. AI brings "speed" and "breadth of net," but it cannot hold "judgment" and "responsibility." The faster drafts appear, the more the process of humans confirming before external release actually gains in weight. Redirect the time saved by speed into the thickness of confirmation ── keeping this balance from tipping is the foundation when bringing AI in.

05Separating Limits from Exaggeration ── AI Itself May Write the Exaggeration

Let us concretely pin down AI's limits in the context of material review. The most dangerous is that AI writes a wrong sentence with exactly the same confidence as a correct one. This is called hallucination (= a plausible-sounding falsehood). Citing a reference that does not exist, mixing up article numbers (writing Article 66 for exaggeration as Article 68, for example), stating that an unapproved use is "common" ── such errors slip into text that looks flawless on the surface.

Another pitfall is that AI becomes not only the detector but also the generator of exaggerated expressions. On request, AI will smoothly write the very phrasing that Article 66 prohibits ── "works well," "superior to other drugs." As long as the tool that creates and the tool that checks are the same AI, humans must stand watch so that the generating side does not slip past the regulatory wall.

06The Key Points Humans Hold ── Four Things Not to Hand to AI

Let us make the unchanging things one degree clearer. The following four are key points that humans keep holding no matter how far AI advances. Let go of them, and the review checkpoint loses its meaning.

What these four share is that each involves value judgment and responsibility. The tasks (create, pick up, tidy) can be handed over, but judgment and responsibility cannot. In the AI era, the reviewer's center of gravity shifts from the one who moves their hands to the one who confirms and signs.

07The Map of All 10 Parts

Let us lay out, in advance, the map of the 10 parts this series covers. Use it as your compass as you read on. With the three functions ── create, detect, verbalize ── as the axis, we proceed while checking the regulatory fence each time.

1. Vol. 01
   AI and Material Review ── What Is Actually Happening (this part)

   The whole-map of the current position and the limits; the starting point of the series
2. Vol. 02
   How to Review Materials That Generative AI Created

   The form of checking, and the pitfalls, in an era when AI writes the draft
3. Vol. 03
   Detecting Exaggeration ── Teaching Article 66 to a Machine

   The mechanism and limits of having AI pick up overstated phrasing
4. Vol. 04
   Checking Citations and Quotations ── Killing Hallucinations at Review

   Preventing nonexistent references and misstated figures on the review side
5. Vol. 05
   Why Judging the Scope of Approval Stays with Humans

   Interpreting indications and dosage/administration, and the line AI cannot reach
6. Vol. 06
   Operating the SIP Guidelines with AI ── Proper Information Provision

   Bringing the measuring stick of the 2018 guidelines down to daily checking
7. Vol. 07
   Learning from the Monitoring Program Report ── Making a Form from Deviation Cases

   Turning MHLW's primary source into learning material for review
8. Vol. 08
   Review Logs and Traceability ── Who Judged What

   How to keep a record of judgments, including AI's involvement
9. Vol. 09
   Introducing AI to the Review Team ── Stages and Division of Roles

   The order of expanding from low-harm areas while confirming as you go
10. Vol. 10
    Integration ── Rooting AI-Era Material Review in the Organization

    Operational rules, division of responsibility, design as an organization

08Connections to Other Chapters ── Reading Alongside AI Programming, AI Marketing, and Material Review

The AI Material Review series connects to the other chapters of this site as follows. Read together, they give a three-dimensional understanding of AI.

• AI Marketing Vol. 5 ── The Balance of Speed and Review ── In an era when materials are mass-generated by AI, how to balance the speed of creating against the speed of reviewing. This series handles the "receiving side" of that
• AI Programming Vol. 1 ── The Basics of Code Generation ── The principle by which AI writes "a plausible continuation," and the hallucinations that arise from it. The underlying idea for Vols. 4 and 5
• The Material Review Series ── The foundation of review practice itself, with AI set aside. Whether the thing created is a material or code, between "can be made" and "can be used" there is review

The era in which AI enters material review has indeed arrived. But that "entering" does not mean taking over the review. Preparing drafts quickly, picking up candidate suspicious expressions, putting the grounds of a judgment into words ── that is as far as AI goes. Create, detect, verbalize ── each stops at a "draft" or a "candidate." Determining whether it is exaggerated, discerning the scope of approval, measuring the balance of safety, and signing off on the decision to release externally: these four remain in human hands.

The conclusion this map points to is simple. Let AI create and pick up; let humans determine and sign. To the exact degree that drafts appear faster, make the process of confirming thicker. Especially in material review, the fences of the PMD Act (exaggeration Article 66, unapproved Article 68, information provision Article 68-2), the Fair Advertising Standards, and the SIP Guidelines apply with exactly the same force to text that AI wrote. Next time, map in hand, we move to the most pressing entrance ── how to review materials that generative AI created.