Give the Rules to the AI ── Designing the Guardrails── Building the Pharmaceutical Affairs Act and the Sales Information Guideline into the AI as a frame, up front

01Rules Have a Hierarchy ── the Upper Binds the Lower

The rules a material must obey are not a single flat list. They form a hierarchy with an up and a down. Without this map first, you cannot decide what to make the AI prioritize. The higher something sits, the stronger it is; anything lower is void if it conflicts with what is above.

What matters is the order. A lower layer cannot overwrite a higher one. Even if a company decides internally "this wording should pass," it will not pass if it breaches the Standards for Fair Advertising. When you give the rules to an AI, you have to teach it this up-and-down relationship as well; otherwise the AI looks only at the internal template in front of it and overlooks the higher-level regulation. Build the fences from the outermost inward. That is the starting point of the design.

02Turning the Act and the Sales Information Guideline into a "Frame" ── Teach the Prohibitions First

Now we make the outermost fence concrete. For the Act's advertising rules, pin down the exact location of the provisions first. The ban on exaggerated advertising is Article 66, the ban on advertising unapproved drugs and the like is Article 68, and the proper conduct of information provision within sales information activities is Article 68-2. Mistake the numbers and the very rule you embed in the AI is wrong, so treat this as a fixed, non-negotiable fact.

The Sales Information Guideline ── properly the "Guideline on Sales Information Provision Activities for Prescription Drugs" (notice of the Director-General of the Pharmaceutical Safety and Environmental Health Bureau, MHLW, 2018) ── is the document that sets out how this information provision activity should be conducted in the field. The key point is that it targets not only the wording of the material but the "activity" itself. It reaches into whom you may tell, on what supporting evidence, and how far.

Here is one design insight. For an AI, teaching "what must not be written" is more effective than teaching "what may be written" first. What may be written is countless, product by product; what must not be written is fixed into categories by the regulation. Stand up the fence of prohibitions first, and the range the AI can generate automatically narrows toward the safe side. A fence is built not to make a path, but to keep you from falling off the cliff.

03Approved Information and Prohibited Terms ── Give It Two Dictionaries

When you bring the frame down to concrete data, what you give the AI comes down to two "dictionaries": the dictionary of approved information and the dictionary of prohibited terms. The former sets the upper bound of "you may say this much"; the latter sets the lower bound of "you must not say this."

The dictionary of approved information is a structured version, kept as-is, of each product's approved indications and efficacy, dosage and administration, warnings and contraindications. As we saw in Vol. 3, efficacy can only be spoken of within the approved scope. To keep the AI from giving in to the temptation to "widen the indication," you hand it the approved wording as the ground-truth data first, so that expressions straying from it can be rejected.

The dictionary of prohibited terms is a list of "words that are dangerous to use," derived from the regulation and the standard. Organized, it breaks down as follows.

These two dictionaries are not things for the AI to "sense on its own." They are things you hand it from outside as explicit data. When the approved information is revised, swap out the dictionary; when a new offending expression is found, add it to the prohibited terms. Write the rules into the code or the prompt and you end up fixing everything every time something changes. Keep them outside as dictionaries and you only have to replace the contents.

04Designing the Prompt and the Check ── Build a Two-Stage Fence

Once the dictionaries are in place, where do you make them take effect? Build the fence in two stages: the fence before generation (the prompt) and the fence after generation (the check). Either one alone leaks.

The first stage embeds the regulatory frame into the instruction given to the AI (＝ the prompt, the command that writes out what you want the AI to do). "You create materials for prescription drugs. The approved indications and efficacy go this far. Do not use the following prohibited terms. Do not touch transaction terms." ── you hand these frames over first, as the premise of generation. This makes dangerous copy less likely to emerge in the first place.

But the first stage alone is not enough. As we saw in Vol. 1, an AI can pretend to obey the instruction while mixing in a plausible deviation. So you place the second stage ── an automatic check after generation. You collate the produced wording against the dictionary of prohibited terms and cross-check it against the range of approved information. Here you do not rely on the AI's judgment; you cut things off with a mechanical match.

This idea of "separating the maker's role from the checker's role" is the same as a founding principle of material review itself. For the same reason a creator does not approve their own material, you do not let the AI self-approve either.

05Rules Go Stale ── Designing for Update and Maintenance

This is the most overlooked point. A guardrail is not done once you build it. Rules go stale. Approved information is revised, the interpretation of the Standards for Fair Advertising is updated, and the operation of the Sales Information Guideline changes too. A frame that was correct last year is a hole this year ── that is the scariest thing.

So at the design stage you decide the maintenance owner and the procedure as well. At a minimum, make it clear whose job each of the following three is.

• When to update ── when there is a partial change to the approval matters, when the standard or a notice is amended. Tie the trigger (＝ the cue for updating) to a concrete event
• Who updates ── assign an owner, such as regulatory affairs for the approved-information dictionary and review for the prohibited-terms dictionary. A rule with no owner rots
• How to record ── keep a version of when, what, and why it changed. Make it possible to trace afterward which version of the rules a past material passed under

Keeping versions has one more meaning. When there is an audit or a challenge, you can show that "this material was reviewed under the version of the rules that was in force at that point in time." Manage the rules as an external dictionary and stamp versions. This runs directly on from the idea, touched on in Vol. 1, of keeping the provenance (traceability) of a generated output.

06The Limits of Automation ── the Fence Does Not Guarantee a "Pass"

We have described how to assemble the fences, but there is a limit that should be stated honestly. Passing the guardrail is not proof that a material is proper. The fence only drops what is "clearly no good"; it cannot play the role of choosing what is "genuinely good."

The reason is that much of the regulation is decided by context. The same word "effective" is fine if used correctly within the approved scope, and becomes exaggeration if used in a groundless setting. The dictionary of prohibited terms leans one way or the other ── either it catches the former too and rejects it, or it misses the latter. Reading context is where today's AI is at its weakest.

So making the fence too strict is also a problem. Swing too far to the safe side and correct expressions get rejected in bulk, burying people in the work of pushing back "this is a false positive." Set the height of the fence by the balance between misses (＝ letting a dangerous line through) and over-detection (＝ stopping a correct line). This is a living parameter, one you keep tuning while in operation.

07Connections to Other Chapters ── the Rules Are the Spine Running Through the Whole Set

The guardrails assembled in this volume connect to the other volumes of this series, and to the sister series, as follows. Rule design is not a standalone technique but the spine running through the whole.

• AI Material Review Vol. 3 ── Efficacy Stays Within the Approved Scope ── this volume's "dictionary of approved information" brings the notion of approved scope, confirmed in Vol. 3, down into a data form the AI can use
• AI Material Review Vol. 5 ── Standardizing Review and AI ── uses this volume's rules as a "common language" that removes the variance between individual reviewers. Share the fence across everyone and dependence on individuals decreases
• AI Programming Vol. 1 ── The Basics of Code Generation ── the idea of inspecting after generation with an automatic check has the same skeleton as verifying AI-written code with tests
• AI Marketing Vol. 1 ── Marketing Redefined ── how to receive mass-generated promotional content within the frame of regulation. The creating side and the reviewing side use the same rules, two sides of one coin

If you are going to have an AI make materials, build the fences before it starts running. That is the core of this volume. Fences from the top down ── the Pharmaceutical Affairs Act as the outer frame, the Standards for Fair Advertising as the yardstick, the Sales Information Guideline as the etiquette, and then the approved information per product. Bring this hierarchy down into two dictionaries, approved information and prohibited terms, and make them take effect in two stages: the prompt before generation and the check after. Separate the maker's role from the checker's role. Rules go stale, so decide up front who updates them and when, and stamp versions.

But do not forget: the fence exists to keep you from falling off the cliff, not to guide you to the right answer. A human judges the pass; the machine stops only the fail. As long as you keep this asymmetry, the guardrail makes review faster and surer. Next time we move on to turning these rules into a common language for every reviewer ── how to reduce dependence on individuals.