Epistemic-Uncertainty

In January 2025, researchers at Mount Sinai hospital tested six leading language models on a simple but crucial medical task: identify fabricated details embedded in patient vignettes. The results were alarming. Across 300 physician-validated cases, hallucination rates ranged from 50% to 82%. DeepSeek’s model hallucinated 82.7% of the time. Even the best performer, GPT-4o, failed half the time. But here’s the truly dangerous part: when these models were wrong, they were more confident. An MIT study from the same month discovered that AI models use phrases like “definitely,” “certainly,” and “without doubt” 34% more often when generating incorrect information than when providing factual answers. ...

The Silent Failure: Why LLMs Can’t Say “I Don’t Know” A patient presents symptoms that could indicate a dozen different conditions. The doctor, instead of saying “I need to run more tests” or “I’m not sure yet,” confidently diagnoses the rarest possibility and prescribes treatment. The patient, trusting the confident delivery, follows the advice. Days later, the condition worsens—not because the original symptoms were untreatable, but because the treatment addressed the wrong disease entirely. ...