01Why It Is Needed ── A Tool Never Exceeds the Skill of Its User
When an AI review-support tool arrives on the floor, hope and anxiety are born at the same time. The hope that "review will get faster," and the anxiety of "is it safe to leave this to AI?" Both are half right. Whether a tool is good or bad is, in truth, decided less by the tool itself than by how well the user understands its quirks.
As we saw in earlier installments, AI quickly puts forward candidate deviations. But whether what it surfaced is truly a deviation, and whether anything was missed — that judgment is made by a person. If the user here lacks literacy, two failures follow. One is overconfidence: because the AI said "no problem," the material is passed. The other is distrust: doubting every one of the AI's flags, ending up reviewing everything by hand, and gaining no speed. Both are the picture of a tool not being handled well.
Literacy is the skill of positioning yourself correctly between these two extremes. Leave to the AI what it is good at, and look yourself at what it is poor at. Only once you can draw that line does the tool help the reviewer. Without that line, even an expensive tool becomes either a machine that merely stamps approval, or one that sits unused and asleep.
02What You Must Know ── The Four Quirks of AI
To use it well, first know the quirks of your counterpart. We have narrowed the properties of AI that a reviewer should grasp at minimum down to four. No difficult theory is required. These are the intuitive markers of "AI makes mistakes in situations like this."
It Writes by Probability
AI does not understand the meaning of a sentence; it merely lines up, by probability, "the word likely to come next in this context." So it optimizes for plausibility, not correctness. However confident it sounds, there is no guarantee of any grounding.
It Lies Without Flinching
Hallucination (= a plausible falsehood). Nonexistent citations, approval content that was never written, mixed-up article numbers — it produces these in exactly the same tone as correct statements. It will not tell you "this is probably wrong."
It Is Pulled by Its Training Data
It is poor at anything absent from the vast text it learned, and gets pulled toward what is abundant there. It leans easily toward the promotional phrasing common in the world, and unless taught, it does not know the latest amendments or your company's own approval information.
It Wavers on the Same Question
Show it the same material twice, and its flags can shift subtly. You cannot expect, as-is, the consistency of a human reviewer. Do not assume a single output is "the one answer."
None of these four is a matter of "AI being at fault." They are properties arising from how it works. Once you understand them as properties, the moments to brace yourself come into view. If a citation is offered, confirm it actually exists (Quirk 02); check the latest approval information yourself (Quirk 03); for important judgments, verify more than once rather than relying on a single pass (Quirk 04) — knowing the quirks turns directly into a verification procedure.
03Understanding the Limits ── What May Be Delegated, and What Must Never Be Handed Over
Once you know the quirks, the limits come into view next. The review work that may be left to AI and the work a person must hold onto are clearly separate. Mix them, and accidents happen. Organized, it looks like this.
| May Be Left to AI ── Preparation | Held by a Person ── Final Judgment |
|---|---|
| Mechanical extraction of prohibited words and superlative expressions | Whether that expression is truly exaggerated within its context (Article 66) |
| Cross-checking approval information against the material's wording, and presenting differences | Whether a difference can be said to exceed the approved scope — the expert line-drawing |
| Flagging missing citations and extracting quoted passages | Evaluating whether that citation is an appropriate basis supporting the claim |
| Ordering, from a large volume of materials, the points needing attention by priority | Reading how patients and healthcare professionals will receive it |
The left column has value in being done fast and without omission, and it is AI's strong suit. The right column is the work of reading context, evaluating meaning, and bearing responsibility — precisely where today's AI is weakest. In particular, "does this phrasing amount to exaggeration" flips black and white depending on context: the same word is fine within the approved scope, yet goes too far when there is no supporting basis. Reading this reversal is, for the time being, the work of a person.
Set the limits correctly, and your stance toward AI is settled too. AI's output is not "the answer" but "preparation." It is a counterpart that lays the ingredients on the cutting board, not the cook who decides the taste. Not letting this positioning collapse is the backbone of literacy.
04The Habit of Verification ── Never Say "Because the AI Said So"
Even when you understand the limits, on a busy floor you tend to accept the output as-is. What prevents that is not willpower but habit. Weave a few confirming actions into every review. Here we narrow it to three.
- Always go to the original source for citations ── Even when AI offers "this paper is the basis," confirm for yourself that the paper exists and actually states that content. Hallucination is most likely to appear right around these citations.
- Check approval information against what is in your own hands ── Efficacy and indications, dosage and administration, warnings and contraindications — cross-check these against the latest approved wording, not the AI's memory. AI's training is always a little out of date.
- The more it says "no problem," the more you doubt ── An "all clear" output is more dangerous than a flagged deviation. An oversight (= letting a dangerous statement through) slips in quietly, in the form of an absent flag. For a "pass" conclusion, form a hunch and re-verify it yourself.
Summed up in a phrase, this habit means never making "because the AI said so" the reason for a judgment. In Vol. 4 we stated that "the role that creates and the role that checks must be separate." Here, a person checks once more the results that the AI checked. Doubling up looks tedious, but for AI-generated materials it is a step that cannot be skipped. Turn the time saved by fast preparation toward this checking — think of it that way, and you will not get the order wrong.
05Preventing Misuse ── Ways of Using It That Must Never Be Done
Literacy includes, as much as knowing the correct ways to use it, knowing the ways that must never be done. Nailing down the prohibitions first is the same thinking used for rule design in Vol. 4. Let us list, by type, the misuses that easily arise on the review floor.
① Self-approval ── Having AI create a material, having the same AI judge whether it is "problem-free," and passing it on that basis. Because the creating habit and the checking habit are the same, they slip past the same holes. The creating party and the checking party must always be separate.
② Deflecting responsibility ── Making "because the AI passed it" the reason for approval. The responsibility for judgment lies with a person in every situation. AI is not a tool for justification.
③ Carrying out confidential information ── Entering unpublished approval-application information or patient data into an external AI service without protection. Draw the line, before using it, on what information may leave the company.
④ Mixing in transactional terms ── Having AI write promotional copy and slipping in transactional inducements such as "a bargain" or "now only." Information provision for prescription drugs and talk of transactions belong to different domains.
The fourth is, rather than something specific to AI, a matter of the boundary of the material itself. Let us confirm it to be safe. What an MR (= Medical Representative), who is responsible for providing information on prescription drugs, may handle is strictly product information provision. Transactions such as price, stock, delivery date, ordering, and discount negotiation are not the MR's domain. Those move between the pharmaceutical wholesaler and the hospital's purchasing department. Have AI write text, and it gets pulled toward the promotional phrasing abundant in its training data, making it easy to cross this line. So the more something was written by AI, the more the reviewer checks with their own eyes that no transactional inducement has slipped in.
What these misuses share is that the wish to take the easy way tries to make AI stand in for judgment and responsibility. AI can stand in for preparation, but it cannot stand in for judgment and responsibility. Not confusing this point is the strongest fence against misuse.
06Cultivation ── From Individual Craft to Organizational Strength
Up to here it has been about a single reviewer. But if AI literacy rests on one person's intuition, the moment that person leaves, the floor can no longer use AI well. Cultivating a personal craft into a shared organizational strength — including this is the design of literacy. Let us organize it in stages.
Share the Quirks
Make the four quirks from Section 2 a shared understanding across the whole review team. Get everyone able to state "AI makes mistakes in situations like this," not just a few knowledgeable people.
Embed Them in the Procedure
Confirming citations, cross-checking approval information, re-confirming "pass" conclusions — write the habits from Section 4 into the review procedure (SOP), not into individual mindfulness. Make the same checks work no matter who performs them.
Make Failures into Teaching Material
Record and share, without hiding, the near-oversights from overtrusting AI and the cases where hallucination was caught. Put them on the previous installment's audit-trail and CAPA machinery, and put them to use in the next review.
Keep Updating
AI's performance, the operation of the Pharmaceuticals and Medical Devices Act, and approval information all change. Last year's common sense becomes this year's hole. Treat literacy not as something acquired once and done, but as a living skill that keeps being updated.
Running beneath these four stages is the idea of making literacy an organizational procedure rather than an individual talent. Do not leave a fine reviewer's intuition untouched as merely something precious. Translate it into a procedure, make it teaching material, and update it. Only then does the skill of using AI well become an organizational asset that is not lost even as people are replaced. This connects straight to the next installment's discussion of governance.
07Connection to Other Chapters ── Literacy Bridges Tool and Structure
This installment's AI literacy is the hinge that bridges the discussion of tools (the earlier volumes) and the discussion of structure (from the next installment onward). Read it together with the related installments as follows, and your understanding becomes three-dimensional.
- AI Material Review Vol. 6 ── The Real Ability and Limits of AI Review-Support Tools ── The installment that looked at the tool's own ability and limits. Receiving that tool with "the skill of the user" is this installment. Tool and user should always be considered as a set.
- AI Material Review Vol. 7 ── Review Records, Audit Trails, and AI ── This installment's "make failures into teaching material" cultivation runs on top of the previous installment's audit-trail and CAPA machinery. Because there is a record, oversights can be put to use next time.
- AI Material Review Vol. 9 ── The Review Structure and Governance of the AI Era ── Beyond cultivating individual literacy into organizational strength lies structure and governance. The next installment advances to that whole design.
- AI Programming Vol. 1 ── The Basics of Code Generation ── The skeleton of a person verifying AI-written code is exactly the same as this installment's person re-checking material that AI checked. On both the creating side and the review side, the responsibility to verify cannot be handed to AI.
An AI review-support tool is not a tool that replaces the reviewer. It is a tool that speeds the reviewer's hands. By exactly as much as work speeds up, the responsibility to verify grows heavier instead. Whether you can take on this asymmetry is the core of AI literacy. Know your counterpart's four quirks; separate the preparation you may delegate from the judgment you must never hand over; do not say "because the AI said so"; and prohibit in advance the ways it must never be used. Then, do not keep that skill in one person's intuition, but cultivate it into an organizational procedure.
No difficult technology is required. What is required is the stance of handling a tool correctly as a tool. AI lines up plausible preparation with astonishing speed. That is precisely why you must not entrust your trust to appearances, and have a person verify meaning and grounding. The next installment advances to the vessel that holds this individual skill ── the whole design of a review structure and governance with AI built in.
- The value of an AI review-support tool is decided less by the tool itself than by "whether the user understands its quirks." Both overconfidence (passing because the AI said no problem) and distrust (doubting everything and ending up with manual work) are the picture of a tool not handled well. Knowing AI's four quirks — writes by probability / lies without flinching (hallucination) / is pulled by its training data and stale / wavers on the same question — turns directly into a verification procedure.
- What may be left to AI is "preparation" such as extracting prohibited words and cross-checking wording. What a person holds is "final judgment" such as whether an expression is truly exaggerated within its context (Article 66), or whether it exceeds the approved scope. The same word is fine within the approved scope, yet goes too far without a basis — reading this contextual reversal is a person's work. Go to the original source for citations, check approval information in your own hands, and doubt "no problem" the most.
- The misuses to avoid are: ① self-approval (having the creating AI do the checking), ② deflecting responsibility (making "the AI passed it" the reason), ③ carrying out confidential information, ④ mixing in transactional terms. The MR is limited to information provision and does not handle transactions such as price, stock, or delivery date — take special care, since AI is pulled toward promotional expressions. Do not keep literacy in individual intuition; cultivate it into organizational strength through the four stages of sharing the quirks, embedding them in the procedure, making failures into teaching material, and updating.
- UNESCO. Guidance for generative AI in education and research. UNESCO, 2023. (International guidance on using generative AI in education and research. Sets out AI literacy and the positioning of human judgment and responsibility.)
- World Health Organization. Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. WHO, 2024. (Ethics and governance guidance on large-scale AI use in healthcare. The warning against overconfidence and the need for human oversight.)
- Ji, Z. et al. Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, Vol. 55, No. 12, 2023. (A review systematically organizing hallucination in generative AI. Why errors in citations and facts occur.)
- Ministry of Health, Labour and Welfare. Act on Securing Quality, Efficacy and Safety of Products Including Pharmaceuticals and Medical Devices (Pharmaceuticals and Medical Devices Act), Articles 66, 68, and 68-2. (The respective articles on the prohibition of exaggerated advertising, the prohibition of advertising pre-approval pharmaceuticals, and the proper conduct of information provision in sales information provision activities.)
- Director-General, Pharmaceutical Safety and Environmental Health Bureau, MHLW. Guidelines on Sales Information Provision Activities for Prescription Drugs. Yakusei-hatsu No. 0925-1, September 25, 2018 (applied April 1, 2019). (The primary source defining the scope, methods, and structure of information provision activities. The basis for the range an MR may handle.)
- Director, Compliance and Narcotics Division, Pharmaceutical Safety and Environmental Health Bureau, MHLW. Explanation of the Standards for Proper Advertising of Drugs and Points to Note. Yakusei-kan-ma-hatsu No. 0929-6, September 29, 2017. (Interpretation of each provision of the Standards for Proper Advertising and operational points to note. Issued by the Director of the Compliance and Narcotics Division.)