People Who Have Never Seen A Patient Assert That AI Will Replace The Ones Who Do
Headlines regularly proclaim that AI chatbots will soon render physicians obsolete. A closer examination of what physicians actually do, and what it would take to automate that work responsibly, tells a more complicated story.
Why autonomous diagnosis and prescribing remains a distant frontier
Even a task as seemingly routine as diagnosing hypertension and adjusting a medication involves layers of clinical judgment: synthesizing a history that spans years, weighing drug interactions against the patient's lived circumstances, assessing whether they can afford or will adhere to the prescribed regimen, reading non-verbal cues, and maintaining the therapeutic relationship that determines whether the patient acts on the guidance at all.
For AI to perform this work autonomously, the field would need regulatory approval frameworks that do not yet exist, liability structures that no insurer has written, and outcome data from randomized trials in real clinical populations rather than curated datasets. The FDA's pathway for Software as a Medical Device remains nascent. State medical practice acts were not written with autonomous AI prescribers in mind. These are not minor technical footnotes. They represent years of policy and legal work at minimum, and that timeline should be taken seriously.
Physicians should be at the table when designing systems meant to do autonomous clinical work. Physicians understand how decisions are made and what patients need. Diagnosing and determining treatments are often more nuanced than "if this diagnosis, then that treatment"
Innovation in healthcare is essential; the stakes demand a different pace
Innovation in medicine is not optional. Stagnation costs lives just as surely as a bad drug or a flawed algorithm. The difference between healthcare and almost every other industry where software disruption has flourished is straightforward: when a social media feed algorithm misfires, users see an irrelevant advertisement. When a clinical decision support tool misfires, a patient may receive the wrong treatment, the wrong dose, or no treatment at all.
The "move fast and break things" ethos that built fortunes in consumer technology is genuinely dangerous when applied to clinical care. Healthcare has learned this lesson before, at significant cost. It pays to think through impacts and proceed with caution, especially when the stakes are high. Some examples:
Theranos promised to revolutionize diagnostics with fingertip blood tests that could run hundreds of lab panels from a single drop. The technology was deployed to real patients before it worked. Physicians made clinical decisions based on inaccurate results. The harm was direct, measurable, and in some cases irreversible. The disruption was real. So were the consequences.
IBM Watson for Oncology was deployed at major cancer centers with considerable fanfare as an AI system capable of recommending treatment plans. Internal documents later revealed that clinicians at the institutions that built it found the recommendations unsafe in a meaningful number of cases. The product was scaled before the evidence base existed to support it, and cancer patients were the ones exposed to that gap.
Sepsis prediction algorithms Multiple health systems deployed proprietary sepsis early-warning algorithms under pressure to improve outcomes. Subsequent independent research found that several of these tools generated high false-positive rates, consuming nursing time on patients who were not septic while potentially desensitizing staff to alerts. An algorithm that cried wolf in a busy ICU is not a neutral error.
The pattern across these examples is consistent. A compelling idea, genuine potential, pressure to scale quickly, insufficient validation in real-world conditions, and patients absorbing the cost of the learning curve. Iteration and safety validation are not bureaucratic obstacles to innovation. They are how medicine earns and maintains the trust that makes adoption possible in the first place.
The most durable health technology innovations, from statins to laparoscopic surgery to mRNA technology, moved through cycles of hypothesis, evidence generation, clinical adoption, and refinement. That process is slower than a product launch. It is also why those interventions are used on hundreds of millions of people with reasonable confidence in their safety profile. Innovation is mandatory, but we must remember our oaths to "do no harm".
Real benefits, and real risks that deserve equal attention
AI-assisted diagnostics are already producing meaningful results in carefully supervised settings. Flagging a missed finding on a radiology image, surfacing a dangerous drug interaction, predicting patient deterioration, and reducing documentation burden through ambient transcription tools are genuine contributions to care quality and clinician well-being.
The risks deserve equal attention. Autonomous systems can fail silently, with high confidence, on atypical presentations. They encode the biases present in their training data, which means historical disparities in care risk being industrialized rather than corrected. When errors occur, accountability is genuinely unclear. Determining responsibility among the vendor, the health system, and the supervising clinician is a question that law, regulation, and ethics have not yet resolved.
Technology cannot substitute for policy reform
The physician shortage in the United States is primarily a product of policy decisions, not clinical inefficiency. Residency training slots have been capped by CMS since 1997. International medical graduates face years of licensure barriers. Rural and underserved communities lack the economic conditions that make sustained practice viable. Administrative burden drives early retirement at scale.
Deploying an AI chatbot to a rural county without a cardiologist addresses none of those root causes. It may create the appearance of progress while deferring the harder work of reform. Substantive solutions include lifting the GME cap to expand residency training, streamlining interstate licensure compacts, creating meaningful loan forgiveness and practice incentives for underserved settings, building sustainable team-based care models, and establishing telehealth reimbursement parity that connects patients to the right specialist regardless of geography.
AI and autonomous systems can support those solutions by reducing the administrative overhead that consumes 30 to 40 percent of a physician's day, by extending reach through asynchronous triage, and by bringing decision support to settings that lack subspecialty depth. The emphasis belongs on support.
Where autonomous AI belongs right now
The most defensible and highest-value deployment of autonomous AI today is in administrative work: prior authorizations, referral correspondence, coding and documentation, scheduling optimization, and care gap outreach. These tasks are rule-bound enough to be performed reliably, they consume an enormous share of physician time, and errors carry lower and more recoverable patient risk. Deployed under human review, this category of automation offers a genuine quality-of-life improvement for the workforce and a measurable cost reduction for the system. Building trust here, demonstrating safety, and earning the right to expand scope incrementally is the responsible path forward.
What patients are actually asking for
Research on patient preferences around AI in healthcare returns a consistent finding: patients want their physician to have more time for them. They want someone to have reviewed their chart before entering the room. They want results communicated in plain language. They want the logistics of care, scheduling, refills, referrals, to be effortless. They are not, by and large, seeking to replace their doctor with a chatbot. They want technology to free their doctor to be more present.
That is the real opportunity on the table. AI that restores the time and attention that made medicine meaningful, rather than systems designed to replace the clinician at the center of the relationship.
The technology is advancing rapidly. The policy work is slow, unglamorous, and essential. Progress requires both, and enthusiasm for one should not become a reason to neglect the other.
A knowledgeable patient is a prepared patient. Learn how the system works to protect your health and your assets. “The Prepared Patient: Your Guide to Surviving the Health Care System” to be released in August 2026 and is available for preorder from Amazon, Barnes and Noble and Johns Hopkins University Press.