IA fondée sur les preuves

Réponse rapide

Evidence-based AI refers to artificial intelligence systems that ground their outputs in verifiable, peer-reviewed evidence rather than relying solely on pattern-learned associations. In clinical contexts, this means every AI-generated recommendation is linked to its original source in the medical literature.

Source : Clinical AI Report, mai 2026

Définition

Evidence-based AI is a design philosophy for clinical AI systems that prioritizes traceability and verifiability. Rather than generating plausible-sounding answers from learned patterns (which risk hallucination), evidence-based AI tools retrieve, synthesize, and cite specific peer-reviewed papers, clinical guidelines, and regulatory documents to support every recommendation. This approach directly addresses the hallucination problem that affects general-purpose LLMs applied to medicine.

Evidence-Based AI vs General-Purpose AI

General-purpose LLMs like GPT-4 or Claude generate text based on statistical patterns learned during training. While often accurate, they can hallucinate — producing confident but fabricated medical claims with no underlying source. Evidence-based AI systems differ by grounding outputs in a retrieval step: the system first searches a curated corpus of peer-reviewed literature, then generates a response that cites specific sources. This retrieval-augmented generation (RAG) approach provides a verifiable evidence chain that physicians can audit.

How Evidence-Based AI Works in Practice

Clinical AI products may retrieve journal articles, guidelines, drug references, or curated topic summaries and attach citations to generated answers. Compare whether the cited source is accessible, current, and relevant, and whether it supports the particular statement. The size of a reference collection and a product’s editorial score do not by themselves establish answer accuracy.

Why Source Transparency Matters

In clinical practice, an unsourced recommendation is an unverifiable one. Evidence-based AI addresses this by making the evidence chain transparent. Physicians can click through to the original PubMed abstract, guideline document, or FDA label to verify the AI's reasoning. This builds clinical trust, supports medicolegal documentation, and enables physicians to weigh the quality of the underlying evidence — rather than treating the AI as an opaque authority.

Mis à jour 7 octobre 2026