I am a hospital-based physician, and despite all the hype, artificial intelligence remains an unpopular subject among my colleagues. Not because we see it as a competitor, but because—at least in its current state—it has proven largely useless in our field. I say “at least for now” because I do believe AI has a role to play in medicine, though more as an adjunct to clinical practice rather than as a replacement for the diagnostician. Unfortunately, many of the executives promoting these technologies exaggerate their value in order to drive sales.
I feel compelled to write this because I am constantly bombarded with headlines proclaiming that AI will soon replace physicians. These stories are often written by well-meaning journalists with limited understanding of how medicine actually works, or by computer scientists and CEOs who have never cared for a patient.
The central flaw, in my opinion, is that AI lacks nuance. Clinical medicine is a tapestry of subtle signals and shifting contexts. A physician’s diagnostic reasoning may pivot in an instant—whether due to a dramatic lab abnormality or something as delicate as a patient’s tone of voice. AI may be able to process large datasets and recognize patterns, but it simply cannot capture the endless constellation of human variables that guide real-world decision making.
Yes, you will find studies claiming AI can match or surpass physicians in diagnostic accuracy. But most of these experiments are conducted by computer scientists using oversimplified vignettes or outdated case material—scenarios that bear little resemblance to the complexity of a live patient encounter.
Take EKGs, for example. A lot of patients admitted to the hospital requires one. EKG machines already use computer algorithms to generate a preliminary interpretation, and these are notoriously inaccurate. That is why both the admitting physician and often a cardiologist must review the tracings themselves. Even a minor movement by the patient during the test can create artifacts that resemble a heart attack or dangerous arrhythmia. I have tested anonymized tracings with AI models like ChatGPT, and the results are no better: the interpretations were frequently wrong, and when challenged, the model would retreat with vague admissions of error.
The same is true for imaging. AI may be trained on billions of images with associated diagnoses, but place that same technology in front of a morbidly obese patient or someone with odd posture and the output is suddenly unreliable. On chest xrays, poor tissue penetration can create images that mimic pneumonia or fluid overload, leading AI astray. Radiologists, of course, know to account for this.
In surgery, I’ve seen glowing references to “robotic surgery.” In reality, most surgical robots are nothing more than precision instruments controlled entirely by the surgeon who remains in the operating room, one of the benefits being that they do not have to scrub in. The robots are tools—not autonomous operators.
Someday, AI may become a powerful diagnostic tool in medicine. But its greatest promise, at least for now, lies not in diagnosis or treatment but in administration: things lim scheduling and billing. As it stands today, its impact on the actual practice of medicine has been minimal.
[link] [comments]