When we say "AI in healthcare", what exactly are we talking about in AI?

Not all AI, in healthcare or elsewhere, plays the same role. Some have a direct medical purpose: helping with screening, diagnosis, or therapeutic decisions. Others are support tools: writing summaries, structuring reports, helping with organisation, or automating administrative tasks.

This distinction is important because the regulatory requirements, expected levels of evidence, and risks are not the same. The HAS precisely distinguishes digital medical devices for professional use, for example for diagnostic aid, and technologies without a direct medical purpose used to reduce administrative burden. (Haute Autorité de Santé)

Behind the word "AI", several technical families coexist.

Symbolic AI relies on rules, decision trees, clinical scores, ontologies, or knowledge graphs. This approach offers major advantages: transparency and traceability. It is possible to understand why a conclusion is produced by the AI, follow its reasoning, and identify its limits. Humans can understand which rules were applied and why a conclusion was proposed.

Machine Learning and Deep Learning learn from data. They are particularly used to recognise patterns in images, signals, or structured records. They can perform very well on a specific task, provided they have representative, well-annotated data - meaning accompanied by a reliable response - that is adapted to the context of use.

Large language models, or LLMs, manipulate natural language. They can summarise a file, rephrase an explanation, extract information from a text, prepare a summary, or help write a report. Their strength is their flexibility; their limitation is that they can produce a convincing response without it being accurate.

RAG, for Retrieval-Augmented Generation, consists of connecting a language model to selected sources — guidelines, protocols, reference standards, patient records, or scientific literature — before producing a response. This does not make the response automatically true, but it reduces the risk of error and improves its factual grounding and verifiability.

Agents orchestrate several steps: querying a document database, applying a rule, extracting a data point, verifying information, and then producing a summary. In healthcare, their interest lies not in "reasoning alone", but in chaining together useful, controllable steps integrated into clinical practice.