Artificial intelligence is revolutionizing medicine, but there’s one problem we can’t ignore: hallucinations. And I’m not talking about patients in a pre-anesthesia delirium—I’m talking about AI models that fabricate diagnoses, symptoms, and therapies, passing them off as absolute truths.
A group of 25 experts from MIT, Harvard Medical School, and Google recently published a study analyzing this phenomenon. AI systems, especially in clinical settings, tend to generate information that sounds plausible, complete with sophisticated medical jargon. The problem? It’s completely false. And the risk of serious errors is right around the corner.
This is a topic I know very well—I teach it in my Cyber-Humanities course at Vita-Salute San Raffaele University in Milan as part of the Medicine and Surgery program. The course is designed for third-year students to help future doctors understand the impact of technology on their work, balancing opportunities with real risks. And that’s the point: artificial intelligence can be an extraordinary ally, but only if those who use it know how to interpret it critically.
Take Whisper, for example—OpenAI’s transcription system used in several hospitals. It’s a valuable tool for converting doctor-patient conversations into digital records, if only it didn’t occasionally invent phrases that were never said. Now, imagine a doctor reading an inaccurate transcription and making a wrong decision. It takes very little to turn a promising technology into a serious problem.
That’s why it’s essential for healthcare professionals to be trained to recognize the limitations of these tools. We simply can’t afford to blindly hand over life-and-death decisions to an algorithm. The future of digital medicine won’t be built on blind trust in AI, but on knowledge, human oversight, and critical awareness.
