From symbolic logic to clinical agents: seven decades of artificial intelligence in medicine — historical-critical review and case description (Article Preprint)

Background. Artificial intelligence (AI) in medicine predates large language models by almost seven decades, a history of recurring promise, failure and partial success that is rarely used to interpret current evidence.Objective. To reconstruct the historical trajectory of medical AI from its conceptual foundations (1943–1959) to October 2026 and to derive from it criteria for appraising contemporary technologies.Methods. Historical-critical narrative review guided by the SANRA scale, organized into five eras, with seven purposively selected cases representing paradigm shifts, implementation failures or evidence milestones.Results. Five eras were identified: logical-probabilistic antecedents (1943–1969); expert systems (1970–1989); the statistical and clinical-informatics turn (1990–2011); deep learning for medical imaging (2012–2021); and foundation models, language models and agents (2022–2026). In every era, performance under controlled conditions preceded, and often exceeded, evidence of patient benefit. In 2026, language models outperform generalist physicians in simulated text-based diagnosis, yet the largest pragmatic primary-care trial found no difference in treatment failure (2.2% vs 2.0%). Robust outcome evidence remains concentrated in imaging-based screening.Conclusion. The central problem of medical AI has shifted from technical capability to clinical validation, workflow integration, governance and the preservation of human expertise. History offers a stable set of questions with which to judge each new technological wave.