How a 10x Insulin Dosage Error Exposed AI Transcription's Confidence Problem

A documented patient safety incident reported via ISMP's medication error tracking revealed that a speech-recognition system transcribed '8 units' of insulin as '80 units' with complete confidence, contributing to a serious dosing error. Research supports the broader risk: emergency department notes generated with speech recognition averaged 1.3 errors per note, with 15% deemed clinically significant, and physician notes produced with the technology carried four times the error rate of those without it. Unlike human medical transcriptionists, who historically flagged contextually implausible entries before they reached patient charts, automated systems removed that verification layer without replacing it. Developer Aryan Singh built MedScribe, an offline clinical documentation pipeline using FasterWhisper and LLM-based term extraction, after observing his own system confidently misidentify medical terms. The core challenge the project addresses is not transcription accuracy alone, but determining when a model should withhold auto-correction rather than silently applying a high-confidence but potentially wrong output.