The Physician's AI Literacy — The Power of the One Who Uses the Tool— What to know, understanding the limits, explaining it to the patient

01Why Literacy Is Needed Now

Unlike an electronic medical record or a diagnostic instrument, AI has a particular property: it "errs plausibly." A calculator that gets a digit wrong produces an obviously strange number, but generative AI (= AI that produces text and images) presents content that is not factual in a confident tone. This is called hallucination (= the phenomenon in which AI fabricates information that has no basis in fact). The more polished the output looks, the harder it is for the recipient to doubt it.

That is exactly why the safety valve, when AI is placed in the clinical setting, lies not on the side of the tool but on the side of the physician using it. The Topol Review (2019), compiled by the UK's NHS, listed education and literacy of the healthcare workforce as the first condition for bringing digital technology into medicine. Grow the people who use it before you bring the technology in — that ordering is the key point.

When only the convenient tool spreads while literacy remains insufficient, three things happen at once.

• Overconfidence — adopting AI's answer without verification, carrying the error straight through to the patient
• Excessive distrust — the opposite: refusing to use it at all, discarding support that could genuinely have been obtained
• The explanatory void — the physician can no longer tell the patient why a given judgment was reached

02What to Know — The Minimum Map

Physicians do not need to be able to derive the equations of machine learning. What they need is to be able to explain the properties and boundaries of the tool in plain words. The World Health Organization (WHO), in its 2021 Ethics and governance of artificial intelligence for health, placed transparency, accountability, and safety as the basic principles of medical AI. Working backward from these, the map a clinician should hold narrows to the following four.

These four points hold in common whether the specialty is internal medicine, surgery, or radiology. The American Medical Association (AMA) deliberately used the term "Augmented Intelligence," positioning AI not as something that replaces the physician but as something that reinforces judgment — the same line of thought. The stance is that the subject of the sentence remains the physician to the very end.

03Understanding the Limits — Can You State What It Cannot Do

The core of literacy is, in fact, not "what AI can do" but being able to state concretely "what AI cannot do." What it can do is taught by the marketing; what it cannot do you have to discern for yourself. The limits that matter most in the clinic are the following three.

The third, in particular, is the domain only a physician can fill. AI learns the tendencies of an average population, but it does not know what the single person in front of you holds dear, or within what kind of daily life they will continue treatment. Even with the same test result, the treatment chosen changes from patient to patient. This "translation from the general to the particular" arises not from data but from dialogue, and it is the physician's own distinctive work.

04The Habit of Verification — Confirm Before You Believe

Literacy is acquired not only as knowledge but as a habit. When you receive AI's output, install into your body the reflex of confirming it. This is the same idea as a pilot who does not take the instruments at face value but cross-checks them. Here are the confirmation steps usable in the field, listed from the heaviest.

1. Go to the source — confirm in the primary sources whether the paper or guideline AI cited actually exists. Do not judge from the summary alone.
2. Cross-check the figures — see whether the dose, test value, or probability contradicts the approved information or the standard criteria.
3. Corroborate by another route — the heavier the judgment, the more you should cross-reference with sources other than AI (textbooks, colleagues, specialists).
4. When in doubt, don't use it — if you cannot fully confirm it, remove that output from your basis for judgment. The courage not to use it is part of the habit too.

05Explaining It to the Patient — Transparency as a Responsibility

When AI is used in care, patients often do not know it is there. The reason WHO's principles list transparency first is to close this asymmetry. What is asked of the physician is not to obscure with jargon but to convey the tool's place in plain words.

An actual explanation holds up fully within a frame like the following.

• That it was used — state the fact: "I also used an AI-based support tool to read the image."
• That it is strictly an aid — show the subject of the sentence: "The final judgment I make, and I take responsibility for it."
• Including its limits — add honestly: "AI is not all-powerful either, so I consider it together with other tests."

This explanation does not stoke the patient's anxiety. Rather, it delivers at once the reassurance that "this is not left to a machine" and the transparency of who the agent of judgment is. Trust is born not from hiding the convenience but from disclosing the limits along with it.

06Designing Education — Literacy Is Not Innate

The literacy described so far is not a talent but something acquired through education. In Japan too, the Ministry of Education, Culture, Sports, Science and Technology's Model Core Curriculum for Medical Education (FY2022 revised edition, 2022) incorporated basic grounding in data science and AI. It is a move to build, from the student years, the foundation for handling the properties and limits of the tool.

Yet pre-graduation education alone is not enough. Because AI changes on the order of months, continuous post-graduation learning is indispensable. Educational designs that function in the field share common elements.

• Use your hands — not lectures alone, but actually using it and deliberately making it err to feel the limits firsthand
• Share failure cases — the case of noticing a hallucination teaches more deeply than a story of what went well
• Keep up with updates — do not treat learning as done once; hold in the organization a mechanism for relearning as the tool changes

For a pharmaceutical company's medical affairs division, this is also a challenge of designing the point of contact. When a physician wants to learn the limits of a tool, where is the neutral information they can rely on? This question is inseparable from the regulatory framework touched on in the next section.

07Connections to Other Chapters — The Fences of Regulation and Information Provision

A physician's AI literacy is back-to-back with how the pharmaceutical side provides information. When AI-produced materials and information reach physicians, there are the fences of the Pharmaceutical and Medical Devices Act and various guidelines. Here we confirm precisely the articles that are easily confused.

• Prohibition of exaggerated advertising — Article 66 of the Act. It prohibits expressions that exaggerate or mislead as to efficacy, effect, or safety. Even if AI generates an "inflated" expression, the regulation does not loosen.
• Prohibition of advertising unapproved drugs — Article 68 of the Act. A provision that unapproved drugs must not be advertised.
• Due care in providing information — Article 68-2 of the Act. It sets a duty of endeavor when providing drug information.

The guideline for interpreting and operating these is the "Guideline for Sales Information Provision Activities for Prescription Drugs" (commonly, HanteiG), issued in 2018 as a notice from the Director-General of the Pharmaceutical Safety and Environmental Health Bureau of the Ministry of Health, Labour and Welfare. In addition, the Standards for Fair Advertising of Drugs — the yardstick for advertising expression — are set out as a notice from the Director of the Compliance and Narcotics Division, Pharmaceutical Safety and Environmental Health Bureau, MHLW. No matter how clever AI becomes, the fact that the information reaching physicians sits within these fences does not change. The physician's literacy and the sender's regulatory compliance are two pillars supporting the same trust from both sides.

It connects to other chapters of this site as follows. The diagnostic-support and literature-summarization installments of the AI Medical series are the premise of this literacy, and the next installment on pharmaceutical companies' AI strategy addresses the design on the sender's side.

AI is not a tool that takes work away from the physician but a tool that tests the physician's judgment. Facing a plausible output — ask its origin, put its limits into words, confirm before adopting, and explain it honestly to the patient. For a physician who can perform this sequence, AI becomes a powerful aid. For a physician who cannot, it becomes a quiet pitfall. What separates the two is not the performance of the tool but the literacy of the one who uses it. And that is grown not by talent but by education and habit. The pharmaceutical side, too, stands in the position of supporting the environment in which physicians grow this power — while keeping to the fences of regulation.