Why Artificial Intelligence Will Never Replace Doctors


As artificial intelligence systems like ChatGPT continue to advance, some wonder if AI could one day replace human doctors. However machine learning expert Steve J. Frank argues that while AI has the potential to transform medicine, it should not replace physicians at the frontlines of patient care. Steve explains why AI will never fully substitute for human doctors.

ChatGPT, the now-ubiquitous artificial intelligence chatbot that generates human-like responses to prompts and queries, recently passed a radiology board-style exam. But should we trust the AI system to start treating patients?

As the demand for medical services increases and physician ranks dwindle, health care providers will have to do more with less. AI is poised to play an increasingly prominent role in caregiving as the technology progresses. While AI systems like ChatGPT exhibit bias, make stunning mistakes and hide those hallucinations under a credible veneer of plausibility, they also continuously evolve, learn from errors and strive to mitigate their shortcomings. With corporate titans across big tech staking their futures on AI, we can expect today’s flaws to diminish quickly.  Still, our relationship with machines has always been complicated, and medicine — caregiving — centers on trust. Can we trust AI over care providers to improve our health?

Trusting AI to Make Diagnoses

AI proponents argue that the technology will eventually earn our trust and maybe even supply the empathy and reassurance we associate with caregivers, but today’s machine-generated empathy can seem sterile and inauthentic. That will likely change, however, as AI sidles into more of our everyday interactions. To that end, a recent study showed medical professionals preferred ChatGPT’s responses to patient questions over those provided by doctors 79% of the time, finding the AI more empathetic as well. Believe it or not, some doctors already use chatbot-written scripts to help them with difficult patient conversations.

Certain medical specialists, such as radiologists, rarely interact with patients, instead sharing their findings with other specialists or primary caregivers. In these specialized roles, the social imperative for human participation is much reduced. Replacing medical specialists with AI could, some suggest, close the widening gap between patient demand and available resources. We regularly learn of new AI systems whose performance equals or exceeds human expertise in tasks such as reviewing mammograms or MRI scans. Yet, relatively few medical AI systems are in widespread use. That’s because matching specialist performance on a test is not the same as performing like a specialist. Real-world performance disappointments — especially over time — have kept medical AI systems closer to the drawing board than the doctor’s office.

AI Health Care Concerns Still Outweigh Challenges Human Doctors Face

AI is a prisoner of its training data — medical images with or without cancer, for example, or health records reflecting whether a hospital patient developed sepsis. A well-designed AI system not only learns from these examples but generalizes beyond them. If the examples are incomplete, omit minority patient populations or skew too heavily toward those with or without the condition, predictions will be unreliable.

Even when the training cohort and the patient population are well-matched at the start, they can drift apart gradually — as demographics evolve — or suddenly. In April 2020, the University of Michigan deactivated its sepsis-alerting AI system as its predictions grew dangerously unreliable. The culprit? The coronavirus pandemic, which fundamentally altered the relationship between fevers and bacterial sepsis.

AI also has practical limits. AI systems are typically trained for mainstream cases because training is expensive. Teaching an AI to spot abnormalities in a chest X-ray requires assembling a large sample of images fully representative of the patient population, then having experts examine and annotate each one. That large investment may pay off for common lung ailments, but training for “long-tail” conditions unhappily pairs the higher costs of obtaining scarce images with lower demand. And there may be many such conditions.

These challenges affect AI to a far greater extent than human clinicians. Where the high cost of training makes AI brittle, humans adapt naturally with experience. Each new addition to a physician’s patient roster will likely reflect the demographics of their everyday practice. If those demographics change, so will the physician’s awareness. Sudden disruptions, like a global pandemic, will demand vigilance and a coordinated response as the unexpected reveals itself. Even better, radiologists are not only alert to unusual conditions in an X-ray but also have access to the patient’s history and symptoms, which may combine to suggest conditions not apparent from any single data point.

But humans also have limits. We miss things, don’t always listen and zone out after staring too long at a screen. Consequently, malignant breast tumors can be difficult to spot in fibrous tissue with similar visual density in a mammogram, and clinician fatigue contributes to missing up to 30% of visible breast cancers overall.  

Computers, on the other hand, are tireless and excel at making fine visual discriminations. ChatGPT can assimilate more medical literature than any human. Without question, AI can help physicians save lives — if we’re careful.  

The Best of Both Worlds

AI assistance is an increasingly essential part of medicine, enhancing our ability to detect and diagnose disease like MRI scans and biomarker blood tests did generations ago. No matter how convincingly a medical chatbot can emulate empathy, though, we cannot allow AI to shoulder past physicians on the front line of patient care — regardless of the economic temptations and impressive performance data. Instead, we must prioritize enhancing diagnostics, treatment precision and patient care by leveraging data-driven insights and augmenting the capabilities of medical professionals with collaborative AI. Our health depends on it.

About the author:

A machine-learning innovator and entrepreneur, Steve Frank is the founder of Med*A-Eye Technologies, which develops AI-driven software that helps doctors spot disease in medical images. Before that, he was a patent lawyer and partner at Morgan, Lewis & Bockius LLC. 


Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum

ChatGPT Answers Patients’ Online Questions Better Than Real Doctors, Study Finds

Doctors, hospitals dream big on AI but move slowly

ChatGPT Passes Radiology Board Exam

Performance of ChatGPT on a Radiology Board-style Examination: Insights into Current Strengths and Limitations

8 Big Problems With OpenAI’s ChatGPT

A Mental Health App Tested ChatGPT on Its Users. The Founder Said Backlash Was Just a Misunderstanding.

A mental health tech company ran an AI experiment on real users. Nothing’s stopping apps from conducting more.

AI can detect signs of depression. Should we let it?

Top Challenges of AI in Healthcare: What Businesses Need to Resolve

Patient-centered Radiology | RadioGraphics (rsna.org)

RSNA 19 Puts Spotlight on Radiologists’ Interaction with Patients – Everything Rad (carestream.com)


Explore more topics
Prolific News
Related news stories