Digital Health and Artificial Intelligence Policy
How ASF governs AI tools in its own work and what ASF standards require of facilities using AI in patient care
Foreword
Artificial intelligence is changing how healthcare is delivered, how documentation is generated, and how quality evidence is produced and analysed. ASF is not exempt from this shift — AI tools are used in drafting, evidence review, and administrative functions at ASF itself. The facilities ASF accredits are increasingly deploying AI in clinical decision support, diagnostic imaging, patient communication, and administrative workflow.
This policy addresses two distinct questions. First: what standards govern the use of AI tools in ASF’s own work — in standards development, in surveyor reporting, and in governance? Second: what will ASF’s published standards require of facilities that deploy AI in clinical care, and on what timeline?
These are different questions requiring different answers. An AI tool used to assist in drafting a standard criterion raises questions of accuracy, attribution, and intellectual ownership. An AI tool used in clinical decision support raises questions of patient safety, liability, and human oversight. This policy addresses both, clearly separated.
1. International Framework
This policy is grounded in:
- EU AI Act (Regulation 2024/1689) — applies to French-registered entities; classifies AI systems in healthcare as high-risk; requires transparency, human oversight, and conformity assessment for high-risk AI systems [1]
- WHO Guidance on Ethics and Governance of Artificial Intelligence for Health (2021) — six principles: human autonomy, human wellbeing and safety, transparency, accountability, equity, sustainability [2]
- OECD Principles on AI (2019) — inclusive growth, human-centred values, transparency, robustness and security, accountability [3]
- ISQua Note on AI in Accreditation (2025) — calls on accreditation bodies to address AI in their standards and to be transparent about AI use in their own operations [4]
- Council of Europe Framework Convention on AI and Human Rights (CETS 225, 2024) — the first binding international treaty on AI; ratification includes France [5]
2. AI in ASF’s Own Operations
2.1 Standards development
AI tools may be used in standards development for: literature search and evidence mapping; drafting assistance for criterion text (subject to human expert review and approval); translation assistance (subject to human translator review); and administrative functions such as comment log management.
AI tools may not be used as the sole or primary author of any criterion, guidance note, or evidence assessment. Every published criterion must be approved by a human Council or Revision Panel member with the relevant domain expertise, who takes personal responsibility for its accuracy and appropriateness. The use of AI assistance in drafting is disclosed in the revision file.
2.2 Surveyor reporting
Surveyors may use AI tools to assist in drafting survey reports — for grammar, structure, and plain language — but every finding and conclusion in a survey report must be based on the surveyor’s own direct observation, document review, and interview. AI tools may not generate findings. The surveyor takes personal professional responsibility for the accuracy and completeness of the report they submit. Where AI assistance is used in report drafting, this is disclosed in the report.
2.3 Governance and administration
AI tools used in ASF governance — for meeting management, document control, or communications — must comply with the ASF Data Protection Policy (ASF-DATAPROTECT-001-v1) and the IT & Data Security Policy. Personal data of Council members, surveyors, and facility contacts may not be processed through AI tools without an appropriate legal basis under GDPR and a completed data protection impact assessment.
The EU AI Act requires high-risk AI systems — which includes clinical decision support — to be transparent and to produce outputs that are sufficiently explainable to allow human oversight. A system that produces a diagnostic recommendation without an explainable basis cannot be meaningfully reviewed by a clinician. ASF’s clinical AI oversight criterion (see above) will specifically require that the facility can demonstrate, to a surveyor on request, that any AI system used in clinical decision support produces an output that includes an explanation of the factors that drove the recommendation, expressed in terms a clinician can evaluate against their own clinical judgment. A “black box” system whose outputs cannot be explained may not be used in clinical decision support in an ASF-accredited facility without a specific documented risk assessment and patient consent.
AI-generated synthetic audio, video, and text (“deepfakes”) present a specific risk in telemedicine — a patient cannot reliably verify that the clinician they are communicating with is the person they appear to be, or that an AI-generated message is genuinely from their healthcare provider. The ASF Telemedicine Standard’s patient identification and platform reliability criteria address this risk at the level of platform verification. The next revision of the Telemedicine Standard will add an explicit criterion requiring facilities to disclose to patients when synthetic or AI-generated content is used in any patient-facing communication, and to ensure that no AI-generated content is used to simulate a specific clinician’s communication without that clinician’s explicit consent and the patient’s awareness.
AI systems used for ICD-10/ICD-11 diagnostic coding, procedure coding, or billing classification represent a high-risk application: coding errors affect insurance claims, epidemiological data, and potentially patient care pathways. Any facility using AI for clinical coding must have a documented human review process in which a qualified clinical coder reviews and approves AI-generated codes before submission. Coding AI must not be used as the only basis for a billing submission. This requirement will be incorporated into the Hospital Standard and Ambulatory Clinic Standard in the next revision cycle.
Patient information booklets, discharge instructions, and education materials generated or drafted using AI tools must be reviewed and approved by a qualified clinician before distribution to patients. AI-generated patient materials that contain clinical information — dosage instructions, contraindications, symptom warning signs — carry a patient safety risk if inaccurate. ASF’s 152 Patient Information Booklets are human-authored and reviewed by clinical experts; any AI assistance in their drafting or updating is subject to the same human expert review requirement as criterion drafting in Section 2.1. This applies to materials produced by accredited facilities as well as by ASF.
3. AI Criteria in ASF Published Standards
The ASF Telemedicine Standard already contains criteria addressing technology platform reliability, security, and data privacy (Standards 1 and 6). These criteria apply to AI-powered telemedicine platforms. The following additional AI-specific commitments apply to all ASF standards:
Any AI system used in clinical decision support — diagnostic imaging AI, sepsis prediction, medication dosing algorithms, triage scoring — must have a documented human oversight process in which a qualified clinician reviews and takes responsibility for every AI-generated recommendation before it influences a clinical decision. The next revision of the Hospital and Ambulatory standards will introduce this as an explicit criterion. Consistent with the EU AI Act classification of clinical decision support as high-risk AI, the oversight documentation must be available to the surveyor on request.
Patients have a right to be informed when an AI system has been used in a decision affecting their care. This applies to diagnostic AI, treatment recommendation systems, and administrative AI that affects appointment scheduling or resource allocation. The criterion will be introduced in the next revision of the Hospital Standard’s patient rights and information section (Standard 2).
AI systems trained on data from high-income, predominantly-male, or ethnically homogeneous populations may produce systematically biased outputs when applied to the populations ASF-accredited facilities serve. Facilities deploying AI in clinical decision support must document the training data source, the validation population, and any known performance disparities across patient subgroups. This criterion will be introduced across all relevant standards in the next revision cycle.
4. Implementation Timeline
References
- European Union. Regulation (EU) 2024/1689 of the European Parliament and of the Council (AI Act). Brussels: EU; 2024.
- World Health Organization. Ethics and Governance of Artificial Intelligence for Health. Geneva: WHO; 2021.
- OECD. Recommendation of the Council on Artificial Intelligence. Paris: OECD; 2019.
- ISQua. ISQua Note on Artificial Intelligence in Healthcare Accreditation. Dublin: ISQua; 2025.
- Council of Europe. Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS 225). Strasbourg: Council of Europe; 2024.
