AI mirage: medical algorithms and the vanishing autopsy

In 2026, artificial intelligence (AI) systems are deployed at scale to support clinical decision-making. Algorithms detect cardiac arrhythmias from ECGs, classify skin lesions from photographs and predict deterioration in critically ill patients. These tools are valuable. However, they share a critical vulnerability: they are trained on labelled datasets where the labels (the diagnoses) derive from clinical assessments recorded in electronic health records (EHRs). The fundamental assumption is that these clinical diagnoses represent the ground truth. This assumption merits examination. In my practice as a forensic pathologist conducting forensic pathology reviews in malpractice litigation, I regularly encounter cases where the clinical diagnosis recorded in the patient’s medical record diverges from the findings at autopsy. These discrepancies raise an uncomfortable question for the emerging field of algorithmic medicine: If AI systems are trained on clinical labels without pathological verification, are they being taught to recognise disease or to replicate diagnostic error? In machine learning, the principle ‘garbage in, garbage out’ is foundational. When a predictive model achieves high agreement with clinician assessments, this concordance reflects the model’s ability to capture clinical patterns, not necessarily its ability to identify true disease. If those clinical patterns include systematic errors, the algorithm will learn those errors as features. This is not a theoretical concern. The diagnostic error literature demonstrates that significant discrepancies between clinical diagnoses and autopsy findings occur with meaningful frequency. Shojania et al conducted a systematic review of studies comparing clinical diagnoses with autopsy findings and found that major discrepancies (errors that, if known during life, would have altered patient management) occur in approximately 8–24% of contemporary hospital autopsies (with a median of 23.5%).1 In a more recent analysis, Newman-Toker and colleagues, using best available …