The World Health Organization (WHO) and the International Telecommunication Union (ITU) published a technical brief in 2025 mapping artificial intelligence used in traditional medicine. The brief calls for good-quality, inclusive data, national policy frameworks and global standards for data quality, interoperability and ethical AI use.

Those requirements turn credibility into a set of questions that can be checked before a clinic or health system adopts a tool. A polished demonstration or a single accuracy figure does not answer whether the software has a defined medical purpose, sound evidence, traceable data, usable records or accountable operators.

1. Define the medical purpose before judging the model

The first question is what the software is intended to do and how its output affects care. A tool that digitizes notes does not create the same risk as one that recommends a diagnosis, ranks treatment options or identifies a medicinal ingredient.

The International Medical Device Regulators Forum (IMDRF) defines software as a medical device by its intended medical purpose and categorizes risk using both the healthcare situation and the significance of the information to a healthcare decision. The framework is not a market authorization and does not replace national rules, but it gives buyers a common way to test whether a vendor's risk claims match the stated use.

A clinician reviews a diagnostic reasoning record on a computer screen (illustrative image)

2. Separate a benchmark score from clinical evidence

A score on examination questions, historical records or a curated image set shows performance on that test. It does not establish how the system behaves with different practitioners, populations, devices or incomplete records.

The IMDRF clinical-evaluation framework divides the evidence into a valid clinical association, analytical validation and clinical validation. In practical terms, a developer should show that the output relates to the target condition, that the software processes inputs correctly and that its use achieves the claimed purpose in the target population and setting.

Pages from historical traditional Chinese medicine texts (illustrative image)

3. Trace the knowledge and training data

Traditional Chinese medicine AI may draw on classical texts, modern clinical records, ingredient databases, images or material generated elsewhere on the internet. A credible disclosure should identify the types and dates of the sources, the populations represented, how experts labeled the material and what the developer excluded.

The 2025 WHO and ITU brief treats data quality and inclusion as governance requirements and documents AI applications across several traditional medicine systems. Evidence from one practice, country or dataset therefore cannot automatically validate a product designed for another.

4. Require an explanation and an audit trail

A clinician needs more than a generated conclusion when an output may affect care. The record should preserve the input, model version, output, confidence or uncertainty information, any source material shown to the user and the identity of the person who accepted or rejected the recommendation.

WHO's 2021 guidance makes transparency, explainability and intelligibility one of six principles for AI in health and separately calls for responsibility and accountability. An explanation interface cannot prove that the underlying reasoning is correct, but versioned records allow reviewers to reconstruct what the system and its users did.

A reviewer examines medical software documentation and regulatory records (illustrative image)

5. Test terminology and interoperability together

Technical connectivity is only one part of exchanging a health record. Two systems can send the same file format while assigning different meanings to symptoms, patterns, ingredients or preparations.

Fast Healthcare Interoperability Resources (FHIR), published by Health Level Seven International, structures electronic health information for exchange and supports terminology bindings and implementation guides. WHO also published international standard terminology for traditional Chinese medicine in 2022 to support consistent concepts and definitions in communication, medical records and technical material.

A vendor should therefore identify both the exchange specification and the terminology set it implements. It should also document how local terms map to that set, how unmapped concepts are handled and whether another organization has tested the interface.

A health information specialist reviews a medical data exchange system (illustrative image)

6. Name the regulator, monitor changes and assign responsibility

Regulatory status must include a country, authority, product version and authorized purpose. A clearance or registration in one market does not establish authorization elsewhere, and a low-risk administrative tool may fall under different rules from software that influences diagnosis or treatment.

Responsibility also continues after launch. WHO's health-AI guidance says developers and users should assess applications continuously and transparently during actual use, while preserving human control over medical decisions. Contracts and operating procedures should identify who reviews model updates, records incidents, can suspend the system and responds when a patient or clinician challenges an output.

A gavel beside medical records represents legal responsibility for healthcare AI (illustrative image)

What the six checks can establish

The six checks do not produce a universal pass mark. They expose whether a product's claims can be tied to a defined purpose, a complete evidence package, documented data, reviewable outputs, interoperable records and named oversight.

A missing answer does not prove that a system is unsafe or ineffective. It does show that a buyer, regulator or clinician lacks information needed to assess the claim, and that gap should remain visible rather than being replaced by a demonstration score.