L'IA peut-elle générer une hypothèse scientifique crédible à partir de données expérimentales brutes ?

Les outils comme FunSearch et AI-co-scientist, sortis en 2024, ont fait émerger des hypothèses novatrices en science des matériaux et en biologie, que les humains ont ensuite vérifiées en laboratoire.

Background

Tools like FunSearch and AI-co-scientist, released in 2024, demonstrated the capacity to surface novel hypotheses in materials science and biology that were subsequently validated through laboratory experiments. Current AI systems leverage machine learning to process and analyze large volumes of raw experimental data, identifying statistical patterns and trends that may elude human observers. This analytical capability underpins efforts to automate hypothesis generation, a process traditionally reliant on domain expertise and contextual understanding. However, the formulation of a scientifically credible hypothesis demands more than pattern recognition — it requires integrating mechanistic insights, theoretical coherence, and empirical plausibility. State-of-the-art systems continue to integrate advances in machine learning, natural language processing, and knowledge representation to better contextualize data-derived patterns and bridge the gap between observation and hypothesis. Despite progress, significant scientific and technical challenges remain in embedding causal reasoning and domain-specific knowledge into AI-driven hypothesis formation. Research emphasizes the iterative co-evolution of AI tools and human expertise, where hypotheses are not merely predicted but critically evaluated and refined through experimental validation.