Demystifying artificial intelligence in health: What health policy-makers need to know

1️⃣ Analytical summary

Context and issues — moving beyond the binary narrative

The document addresses a concrete situation: AI (from classic statistical models to generative AIs) is becoming established in care, public health, research, and management, while two opposing narratives dominate — the panacea on one side, the catastrophe on the other (p. xv-xvi). It is aimed at readers of very varied levels, from the novice to the experienced professional. The central problem identified: many promises are overstated, driven by commercial interests and not supported by long-term clinical evidence (p. xix); actual adoption remains slow and mainly confined to low-stakes tasks (transcription, summarisation), while clinical uses remain at the pilot stage and require human supervision (p. xvi).

Operational contributions — literacy, key questions and policy options

The document provides three actionable items: a foundation of definitions and concepts to "demystify" AI (executive summary, p. xxvi-xxxiii); a mapping of health applications classified into four areas — care, public health, research, management — with their degree of maturity (chap. 2, p. 37-39); and above all, a series of structuring questions for decision-makers (chap. 3) as well as a "menu" of policy options broken down by actor (governments, institutions, professionals, patients) in chapter 4 (p. 177-186). The common thread, reusable in training: AI is a means, not an end, and its value depends on the skills and intentions of those who use it (p. xvi).

2️⃣ Key points of the document

1️⃣ AI is not a unique technology — It covers classical machine learning, deep learning, and generative AI, each with distinct uses and risks; public attention focuses on LLMs (such as ChatGPT, Gemini) but the majority of mature health uses involve other techniques (p. xv, p. 5-14).

2️⃣ The "hallucinations" are a structural risk in health contexts — An LLM predicts the most probable word without understanding the truth; it can produce plausible but false outputs, particularly dangerous in care and public communication (p. 17-19). Human supervision and techniques such as RAG reduce — without eliminating — this risk.

3️⃣ Bias and fairness: an issue throughout the life cycle — Non-representative training data can amplify discrimination against already underserved populations (Fig. 3.3, p. 166). The case of people with disabilities illustrates concrete exclusions (Box 3.10, p. 168).

4️⃣ Human supervision and accountability remain central — The principle of "meaningful human control" must guide any high-stakes use; opague systems create a "responsibility deficit" that undermines access to justice in the event of harm (p. 148-157, p. 181).

5️⃣ A menu of options by actor, not prescriptive recommendations — Chapter 4 outlines pathways for governments, institutions, professionals, and patients, around two axes: system preparation / strategic integration, and governance / ethics / rights (p. 177-186). It emphasises: professionals do not need to become experts in AI, but must understand its limitations (p. 178).

3️⃣ Action points for local actors

Calibration reserve:

the document does not provide an operational field checklist. The points below translate, for a local actor, the posture and the policy options described (especially chap. 3-4). They pertain to literacy, vigilance, and participation, not to a technical deployment.

1️⃣ Build basic AI literacy within the team — Take ownership of the foundational definitions and concepts (executive summary, p. xxvi-xxxiii) to distinguish between classical AI, deep learning, and generative AI. Objective: to engage in dialogue with publishers and identify overstated promises (p. xix).

2️⃣ Adopt a question framework before any AI tool — Revisit the questions from chapter 3 (p. 124-176): what is the real purpose? what evidence of effectiveness? is the system evaluated and approved? what human supervision? what mechanism in case of error? The professional does not need to become an expert, but to know how to question (p. 178, p. 184).

3️⃣ Demand human supervision and a right to report — Formalise, for any high-stakes use, the principle of meaningful human control (p. 181) and provide reporting/alert mechanisms — the document calls for protecting concerned whistleblowers from risky AI use (p. 184).

4️⃣ Place equity and inclusion at the heart of use — Ensure that tools are designed for the diversity of audiences (language, culture, disability); rely on the lessons from Box 3.10 (disability, p. 168) and on the requirement for representative datasets (p. 183, p. 185-186) to avoid exacerbating access inequalities.

5️⃣ Inform people and secure their data — Make visible the use of AI in the process (consent, transparency on data), in accordance with the “patient ” options (p. 185). Support secure environments for the processing of health data and GDPR compliance (p. 184).

6️⃣ Open the participation of the public in governance — Draw inspiration from participatory models (citizen panels, co-design workshops): the example of the Belgian citizen panel on AI (Box 3.11, p. 174) shows that a non-expert public can formulate useful ethical guidelines. Unmet need in the document: ready-to-use participatory protocols still need to be built locally.

4️⃣ Additional references

① HAS & CNIL — “Supporting the proper use of AI systems in a care context” (February 2026).

Practical French guide (twelve sheets, from acquisition to decommissioning + governance and generative AI). Complementarity: translates the general framework of the document into obligations and concrete best practices.

⚠️ Status: project submitted for public consultation until 16 April 2026 — the final version may evolve; to be re-verified before publication. Page HAS

② WHO-Europe — “Artificial intelligence is reshaping health systems: state of readiness across the WHO European Region” (Nov. 2025).

First assessment of the integration of AI into the health systems of 50 member states (survey 2024-2025). Complementarity: provides the empirical field data that the primer, conceptual, does not provide. Publication WHO-Europe

③ European Observatory — « Artificial intelligence in public health: lessons from the EPH Conference ».

Article targeting AI in public health (promotion, surveillance, prevention, equity) — the angle closest to the PES field. Complementarity: refocuses the discussion on prevention/promotion rather than on care.

In Health Practices 

Digital technology and artificial intelligence - https://www.pratiquesensante.com/2-6-intelligence-artificielle-numerique

and Learning with, learning digital technology 

5️⃣ Frequently asked questions (FAQ)

1️⃣ « Does “AI” mean the same thing everywhere in this document?

No. The document emphasises that AI encompasses several families (statistics/rules, machine learning, deep learning, generative AI), with different maturities and risks (p. 5-14, p. xxvi-xxxiii).

2️⃣ Can we trust AI in health?

Trust must be earned and calibrated: neither blind faith nor excessive scepticism. It combines a rational assessment of the reliability and the experience of users, and assumes a clear communication about what the tool can and cannot do (p. 15-16, p. xvii).

3️⃣ What is a “hallucination” and why is it serious?

It is a plausible but false output generated by a model that optimises linguistic probability, not truth. In health, justice or public service, this can cause real harm and erode trust (p. 17-19).

4️⃣ Where is AI actually used today?

Especially for low-stakes tasks (transcription, summarisation, administrative tasks); clinical uses remain largely at the pilot stage and require human supervision (p. xvi, chap. 2, p. 91-102).

5️⃣ What is the contribution of AI to public health?

Epidemic forecasting, health behaviour analysis, risk segmentation, improvement of communication — provided there is quality data, interdisciplinary collaboration, and public trust (p. 64-82).

6️⃣ What can a patient / concerned person do?

Be informed about the use of AI, understand what data is processed, be able to opt out of certain uses, and participate in governance through citizen panels, consultations, and co-design (p. 185-186, Box 3.11 p. 174).

7️⃣ Do I need to become an AI expert to be concerned?

No. The document is clear: professionals do not need to become experts, but should acquire a practical understanding of the opportunities and limitations (biases, hallucinations) to make informed decisions (p. 178).

6️⃣ Rewriting in Easy to Read and Understand (FALC)

What is this document about?

• Artificial intelligence (AI) is a set of computer tools.

• These tools learn from data. They perform tasks on their own.

• There’s a lot of talk about AI in health. Some say it is magical. Others are afraid.

• This document helps to understand AI. It avoids exaggerations.

What needs to be remembered

• AI is not a single technology. There are several types.

• Sometimes, AI makes mistakes. It can give a false answer that seems true.

• In health, a mistake can be serious. A human must always check.

• AI can be unfair. It can mistreat certain people.

• We must protect people's data. We must say when we use AI.

• People can give their opinion on AI. Their opinion is important.

• You don’t need to be an expert. You just need to ask the right questions.

7️⃣ Cross-sectional analysis — Values of Health Practices

Literacy : central objective: the document aims to "demystify" AI for readers of varying levels, with a dedicated glossary (p. xxvi-xxxiii) — but remains an expert text, without a simplified version for vulnerable audiences.

Empowerment : patients are described as actors (information, consent, ability to influence governance), not just recipients of care (p. 185-186).

Participation : co-construction is explicitly valued (citizen panels, consultations, co-design workshops; Box 3.11 p. 174), while alerting to the risk of "token" participation (p. 176).

Community health: the collective/population dimension is present through public health (forecasting, behaviours, communication) and community engagement for "contextually appropriate" tools (p. 180).

Ethics : cultural and social biases are identified as a major issue throughout the life cycle (Fig. 3.3 p. 166); WHO ethical principles (Box 3.5 p. 149) and generative AI (Box 3.6 p. 150).

Human rights: equity, inclusion and human rights are set as guiding principles of AI policies (p. 180-181), with attention to historically discriminated populations.

Intersectorality : AI in health goes beyond the health field alone (regulation, freedoms, environment, energy) and calls for interdisciplinary collaboration (p. xx, p. 184).

Partnership : inter-organisational collaboration and co-development with third parties are recommended, with vigilance on intellectual property and supplier dependence (p. 182).

Fight against discrimination: the document explicitly addresses algorithmic discrimination (disability, minorities, underserved populations) and calls for equity audits and data representativeness (Box 3.8 p. 160, Box 3.10 p. 168, p. 185-186).

8️⃣ Assessment of the reliability of the resource

Scientific relevance — high. Institutional publication under the auspices of WHO-Europe / European Observatory, recognised academic and hospital authors, open licence, abundant bibliography (ref. p. 193 and following). The authors practice epistemic caution in line with the PES framework: they remind that most of the evidence comes from pilot studies and not from common usage (p. xxi-xxii), and distinguish hype and reality. Caveat: data 2024-2025 in a rapidly evolving field; some economic/environmental estimates are presented as issues to measure, not as results.

Operational relevance — average for the PES field. The document is a guide to literacy and policy decision-making, not a field tool. It offers neither a checklist nor a protocol directly applicable by a local prevention/promotion actor. Its operational value lies in the critical stance, the common vocabulary, and the structuring questions it allows to address (chap. 3-4).

9️⃣ Strategic hashtags

#HealthPractices #AIinHealth #DigitalLiteracy #PublicHealth #EthicsAI #HealthEquity #DataGovernance #EvidenceBasedPrevention