AI and Healthcare: An Introduction

What the French National Authority for Health’s Recommendations Reveal About the Future of AI in Medical Practice

Since 2004, France has had an independent public authority called the Haute Autorité de la Santé (HAS), whose objectives include evaluating health products, recommending best practices, and improving the quality of care. Here, “health” is understood as “a state of complete physical, mental, and social well-being, and not merely the absence of disease or infirmity.”

To achieve these goals, the HAS regularly publishes strategic plans. The latest covers the period 2025–2030 and identifies several challenges, including:

• “Significant demographic and epidemiological changes”
• “Major social and territorial inequalities”
• “Systemic difficulties affecting healthcare and support services”
• “A major challenge in the financial and environmental sustainability of the healthcare system”
• Changes in “people’s expectations of the healthcare system”

The HAS plans to rely partly on “health innovations” to address these developments. Among these advancements, generative artificial intelligence is recognized as potentially playing a key role. In October 2025, the HAS published a usage guide summarizing current use cases and suggested precautions for users.

For context, in healthcare, AI applications range widely — from drug discovery and administrative management to robotics and pathology identification.

First, it’s worth noting that the scope of AI considered in this guide, intended for healthcare professionals, remains relatively narrow. It includes non-medical solutions like translation or transcription services, as well as tools closer to core medical tasks, such as those that “clarify medico-social pathways.” However, it does not address more direct solutions like diagnostic or treatment planning software, despite past enthusiasm for such applications. For example, in 2018, an article in Annals of Oncology claimed that a convolutional neural network model outperformed a panel of doctors in detecting cancers.

In defense of the HAS, the lack of a clear definition of what constitutes generative AI makes it difficult to distinguish from other forms of AI. However, it’s hard to deny that this reflects a certain level of skepticism toward these technologies, which is directly evident in a second dimension.

To guide the use of AI, the HAS encourages healthcare professionals to work with the technology following the acronym A.V.E.C. (The word “Avec” being “With” in French. Yes, they’re insistent). This stands for:

• Appendre (Learn): Healthcare professionals are encouraged to understand the technical workings of AI and master its use, particularly through “specific continuing education” or by participating in “exchanges organized by universities, national professional councils, professional organizations, or learned societies.” They should also favor systems that allow them to “exchange with the entity responsible for the system” and choose systems aligned with their workflow.
• Vérifier (Verify): Professionals must ensure the AI system’s “compliance with regulatory requirements for its intended use.” They should avoid transmitting confidential information and confirm that recommendations are based on serious, existing documents. They must treat AI output as a “proposal” and “rephrase generated text as needed,” all while “maintaining their independence and professional skills.”
• Estimer (Assess): Medical staff should establish quantitative or qualitative indicators to measure changes brought by the system and analyze its “suitability for their available resources and time.” If working in a large institution, they are encouraged to “engage with decision-making bodies” to “evaluate the long-term relevance of uses and their organizational impacts.”
• Communiquer (Communicate): Professionals must “exchange in appropriate language with the patient or accompanied person about the use of a generative AI system, to enhance understanding and build trust.” They should also “report major errors and limitations of generated content to the provider.”

This paradigm reveals something fundamental: in healthcare, we have not yet reached a stage where delegating responsibility to machines is seriously considered by authorities — even though, for years, machines have demonstrated competence at least equal to humans in certain diseases.

According to a study by the Hub France IA interest group, the market does not currently reflect this dynamic: about 48% of healthcare companies offer services in “medical or therapeutic decision support,” while 44% focus on “medical diagnosis support.” Meanwhile, healthcare professionals, according to the same study, primarily expect systems for “administrative assistance” (51%) or “document management.” Their main concerns are data security risks (62%) and reliability/trust in performance (57%).

This example invites reflection on the role we want AI to play in our societies. Beyond healthcare — where lives are at stake — it’s crucial to consider the level of responsibility we are willing to accept or delegate in the name of performance. It should also prompt us to think about how to ensure we always retain this choice. Faced with the list of challenges outlined by the HAS, perhaps automation will become inevitable. Things will surely evolve over time :)

Louis Fourneau, student at 42 Paris, graduate of Sciences Po Paris

Article written in French without AI assistance, translated in English with the help of Mistral AI.