Medical AI Today — The Whole Picture, and Separating Out the Overstatement— We first draw the types, the current state of implementation, the limits, and the regulatory frame

01What "Medical AI" Is — Taking the Word Apart First

The single sentence "we introduced medical AI" actually blends three separate meanings. Does it help make a diagnosis, predict the future, or take over clerical work? These three differ in how they are used at the point of care, and in how approval and regulation apply. Let's first define, strictly, the terms this article uses throughout.

What this series calls "medical AI" is the umbrella term for systems built around machine learning (= a computational method that learns patterns from data) that support one of the following in medicine: judgment, prediction, or work. One distinction here is decisive.

• AI as a medical device — Software used directly in diagnostic or treatment decisions is, in Japan, classified under the Pharmaceuticals and Medical Devices Act (PMD Act) as a Software as a Medical Device (SaMD = software that functions as a medical device in its own right), and is subject to approval or certification.
• AI that is not a medical device — Operational support such as appointment intake, transcription, or literature summarization, which does not itself enter into diagnostic or treatment decisions, most often does not qualify as SaMD. The same "AI," in regulatory terms, is treated in an entirely different way.

This dividing line is the starting point for everything that follows. If you treat "AI finds the disease" and "AI writes the meeting minutes" as one and the same story, both expectations and the understanding of regulation go astray.

02The Three Types — Diagnosis / Prediction / Operations

We divide medical AI into three by its role at the point of care. This classification also becomes the skeleton of the later installments.

These three types are not independent. A prediction model's output becomes the entry point to a diagnosis, and the data gathered by operational support becomes the raw material for prediction. Yet "how you confirm the effect" and "who bears responsibility" differ by type. Lumping them together as "medical AI is effective," while conflating these, is the entry point to overstatement.

03The Current State of Implementation — How Far Into the Field It Has Gone

As of 2026, the three types are each at a different stage. To see their real capability correctly, let's organize them stage by stage.

The further a domain has advanced in implementation, the more important it is not to miss that it has stopped at "taking over part of the physician's work," not "replacing the physician". Even in approved diagnostic support, responsibility rests with the physician. This structure is the central theme of the next installment (clinical decision support).

04Separating Limits From Overstatement — Questioning What "It Works" Contains

The place people most often stumble in evaluating medical AI is the content behind the words "high accuracy". The same "accuracy" means something entirely different depending on how it is measured. When people in pharma handle it in promotional materials or internal evaluation, they must always separate out the following four points.

• Sensitivity and specificity — "The power not to miss (sensitivity)" and "the power not to mistake (specificity)" are different things. Being high in only one makes it impractical. A claim that shows only one number deserves caution.
• The gap between training data and the field — There is no guarantee that a model trained on one hospital's data delivers the same performance at another hospital with a different patient population. "AUC 0.95 at Hospital X" is not your own institution's performance.
• Retrospective versus prospective studies — Results measured on past data and results measured by building the tool into actual practice are different. The former alone cannot establish that something is "clinically effective."
• The kind of outcome — "Detection rate went up" and "patient prognosis improved" are different. Claiming the latter on the grounds of the former is a leap.

05The Regulatory Frame — Software as a Medical Device (SaMD)

AI used in diagnostic or treatment decisions is, in Japan, treated under the PMD Act as Software as a Medical Device (SaMD), and receives approval or certification after a risk-based class classification. Internationally, too, the IMDRF (International Medical Device Regulators Forum) has organized the definition and risk framework of SaMD, and national systems are aligning with it.

Here lies one difficulty that sets it apart from conventional medical devices. AI changes performance when its training data is updated — that is, "performance at the time of approval" and "performance after update" are not necessarily the same. Regulators in every country are grappling with this question of "how to regulate software that keeps changing."

• Japan (PMDA) — It has organized its thinking on the review of SaMD and operates a framework that anticipates post-market performance change (a system called IDATEN, which permits planned performance improvement after approval).
• United States (FDA) — In 2021 it published the "AI/ML-Based SaMD Action Plan," signaling a direction in which model updates within a "Predetermined Change Control Plan" are permitted without re-approval.

The shared idea is not "approve the finished product once and be done," but "on the premise that it keeps changing, make the manner of change promised in advance". When the pharma side handles an AI product, whether such an "update management plan" exists is a gauge of its reliability.

06How to Think About Safety — Data, Explanation, Responsibility

The safety of medical AI cannot be measured by performance numbers alone. In actual operation, you have to think in the following three layers.

• Data safety — The handling of patient data must follow frameworks such as the Guidelines for the Safety Management of Medical Information Systems (MHLW). In both training and inference, the protection of personal information and the responsibility for its management cannot be set aside.
• Explanation of the judgment — Can the clinician explain to the patient why the output came out that way? An operation that pushes through with "the AI said so" while the internals stay invisible does not pass, either clinically or in regulatory terms.
• Where responsibility lies — When AI produces a wrong output, who bears responsibility? Under the current framework, the final judgment and accountability remain on the clinician's side. AI does not become a bearer of responsibility.

The WHO, too, in its 2021 "Ethics and governance of artificial intelligence for health," lists human oversight, transparency, and accountability as core principles of medical AI. Understand that the performance of the technology and the safety of its operation are things to be confirmed separately.

07The Map of All 10 Installments

We lay out the ten installments this series covers along the three types (diagnosis / prediction / operations) and the axis of regulation and ethics. Use it as your compass as you read on.

1. Vol. 01
   Medical AI Today — The Whole Picture, and Separating Out the Overstatement (this piece)

   Presenting the three types, the current state of implementation, the frame of regulation and safety
2. Vol. 02
   Clinical Decision Support — Designed to Assist, Not Replace, Judgment

   The central issues of diagnosis and prediction; designing so final responsibility stays with people
3. Vol. 03
   Image-Diagnosis AI — The Current State in Radiology, Pathology, and Endoscopy

   The representative domain of diagnostic support; how to read sensitivity and specificity
4. Vol. 04
   Prediction Models — Reading Deterioration and Readmission Ahead of Time

   A deep dive into the prediction type; performance gaps between facilities and the wall to rollout
5. Vol. 05
   Generative AI and Operational Support — Documentation, Summarization, First-Line Handling

   The main battleground of the operational type; managing hallucination
6. Vol. 06
   The Regulation of Software as a Medical Device (SaMD)

   The frame of approval and certification; how to manage updates
7. Vol. 07
   Medical Data and Privacy

   Data protection in training and inference; the safety-management guidelines
8. Vol. 08
   Drug Discovery, Clinical Development, and AI

   Use in exploration, trial design, and analysis; how regulators receive it
9. Vol. 09
   The Ethics of Medical AI — Accountability and Fairness

   Human oversight, bias, explanation to the patient
10. Vol. 10
    Integration — How to Choose and How to Evaluate Medical AI

    The series' conclusion; a checklist for the adoption decision

08Connections to the Other Chapters on This Site

The AI Medical series connects to the other chapters of this site as follows. Reading them together lets you understand regulation, implementation, and ethics in three dimensions.

• AI Marketing series — The promotion of AI-embedded products and services, and the boundary of the PMD Act and the Sales Information Provision Guidelines. The side that handles the "selling" of medical AI.
• Material Review series — With what eye to check materials that involve AI. The craft of separating out the exaggerated and the unapproved.
• Advertising Regulation — The Frame of the PMD Act — The origin of exaggeration (Article 66) / unapproved (Article 68) / the duty of information provision (Article 68-2).

"Medical AI" holds, under one word, three jobs different in nature — helping to diagnose, reading ahead, cutting the workload. These three differ in how their effect is confirmed, in how regulation applies, and in how responsibility remains. What this installment most wanted to convey was the stance of taking the grand single sentence "AI will change medicine" apart and looking at it. Take it apart, and you begin to see how far is already implemented and where the overstatement begins. Diagnostic support, even when approved, stays at assisting the physician; prediction wobbles in performance once it crosses facilities; operational support spreads fast but does not enter into diagnosis in return — with this resolution, you become able to judge for yourself whether a single sentence in a material touches Article 66 or Article 68.

The next installment steps into the center of this map — "clinical decision support," which straddles diagnosis and prediction. How should you design so that it assists judgment while leaving final responsibility with people? We draw the field operation of approved tools and the line-drawing of responsibility from concrete examples.