The Trump administration announced its support of an experiment in Utah where AI is writing prescriptions, plans to offer over $50 million in research awards to developers of conversational AI for cardiovascular care and has created an expedited approval process for digital health products like AI chatbots June 4.
The administration is also working on a regulatory pathway for independent AI physicians, likening the process to the decades-long effort to bring self-driving cars from test tracks to public roads.
Seven physicians recently joined Becker’s to discuss their thoughts on what the limits on AI in healthcare should be and where they think the technology is headed next.
Editor’s note: Responses have been lightly edited for clarity and length.
Question: The White House recently backed a push for AI to take over more of the duties of physicians, including writing prescriptions and conversing with patients. What is your reaction to this news? What sort of things should AI never be allowed to do in healthcare?
Chacko Alappatt, MD. Board-Certified Rheumatologist. Medical Director, Arthritis Osteoporosis Center of Southwest Ohio (Miamisburg). I am optimistic about AI’s role in healthcare, but I don’t believe it should replace the physician-patient relationship, which remains the foundation of trust and effective care. AI is most valuable when it removes administrative burdens and busy work, allowing physicians to spend more time with patients and apply their clinical experience where it matters most. In my experience, patients who use AI often arrive better informed and prepared for meaningful discussions. While AI can rapidly summarize published information, it cannot replace the judgment that comes from years of caring for patients and seeing how treatments perform in the real world. The future is AI-enabled physicians, not AI-directed healthcare.
Eric Lee. Vice President and Chief Health Information Officer. AltaMed Health Services:
Pushing for independent AI physicians is bold yet unrealistic, given where we are at today. Using generative AI as we all know it is only several years old, and is akin to being in its toddler years. Would you trust a toddler or technology in its toddler phase completely? AI has rapidly evolved, improved and delighted in ways we could not envisioned, but it also continues to surprise and befuddle us with the types of errors, including multiple deaths for those who placed complete trust in AI, made in ways we could not have predicted.
Asking for AI to assist with the increasingly onerous administrative burden is reasonable, but wouldn’t the better approach be to go upstream and address why the administrative burden has worsened considerably?
At some point in the next decade or two, I believe we will have semi-independent AI “providers” that will see patients, but need to be closely supervised by a physician not unlike what occurs in teaching programs nationally. The reason we get to this point is not because patients feel comfortable with this concept, but because healthcare will have slowly developed the trust to get to this state. The most important point to emphasize here is when patient safety is at stake, you cannot automate care with an independent AI physician. It must be augmented care with a human provider in the loop.
Alexander Levit, MD. Medical Director of Hospital at Home at Lee Health System (Fort Myers, Fla.): It is important to distinguish what portions of healthcare delivery are being enhanced by AI. Clinical enhancements should be distinguished from operational enhancements. The latter needs to be pursued aggressively as AI will be able to vastly decrease the knowledge costs involved with linking patient acuity to treatment end points. Clinical tasks, including some that do involve tasks behind licensure requirements, should be pursued, albeit more cautiously. Addressing specifically the question of whether an AI agent should be allowed to write refill prescriptions, I see a viable path for this, but the particular case put forward is concerning.
The list of medications in this program, the first of its kind, is lengthy. When sacubitril-valsartan is among the list of medications in the first iteration, this gives me pause. This medication is a great medication for certain types of heart failure; but notably, a sizable number of patients go on to not tolerate it due to both kidney function effects and low blood pressure. It is noteworthy that the proposal from Doctronic puts forth a “Comprehensive Medical Assessment” that is done at the time of a refill. However, the assessment does not mention laboratory monitoring which is necessary to monitor kidney function. Perhaps the LLMs are good enough at escalation when this is in question, but it is unclear.
Second, it does not appear that a physician or provider is involved in the process of informed consent in making the decision to initiate the AI-driven refills. Again, such a process might be acceptable with a lower risk formulary.
Brandon Ortega, MD, an orthopedic spine surgeon at Long Beach (Calif.) Lakewood Orthopaedic Institute: On the White House push: I’m not opposed to AI expanding its role in healthcare, but I think we need to be precise about what we mean. There’s a wide spectrum between “AI drafts a refill for a stable chronic condition” and “AI converses with a patient about new neurological symptoms.” Conflating those is where the policy conversation goes sideways.
In spine surgery specifically, the stakes of a missed finding are enormous, cauda equina syndrome, cord compression, malignancy masquerading as mechanical pain. These are diagnoses where the difference between a good outcome and a catastrophic one is measured in hours. AI doesn’t yet have the clinical pattern recognition to safely gatekeep those presentations, and more importantly, it can’t be held accountable when it gets it wrong.
What AI should never do in healthcare: serve as the final clinical authority on any decision with irreversible consequences, operate without a named physician who owns the outcome or interact with patients in a way that obscures the absence of a licensed clinician in the loop.
What AI does well, and what I use it for, is everything that isn’t that. Documentation, research synthesis, administrative workflows, surgical planning support. It makes good physicians faster. That’s the use case worth investing in.
Leonid Pravoverov, MD. Nephrology Specialist at Oakland (Calif.) Medical Center: I think the discussion is often framed too much as AI versus physicians. The more important question is how technology can help us deliver safer, more coordinated and more personalized care. In integrated systems with mature digital infrastructure, AI is a natural extension of existing clinical workflows and decision-support tools rather than a replacement for clinicians.
I support thoughtful adoption of AI to reduce administrative burden, improve care coordination, identify care gaps, and support clinical decision-making. However, accountability for diagnosis, treatment decisions, and prescribing medications should remain with licensed clinicians. AI should never independently make irreversible medical decisions, determine access to care without transparency and oversight, or replace the physician-patient relationship in situations requiring judgment, empathy and shared decision-making.
Easwar Sundaram Jr., MD. President of Texas Institute for Neurological Disorders (Dallas): While AI interaction will make some tasks better, making clinical decisions and writing prescriptions cannot be a machine learned option but a decision based on personal interaction with the patient and considering a multitude of options. Automating several redundant processes is welcome but not core physician duties.
Fred Watkins. Plastic Surgeon at Inova Fairfax Hospital (Rockville, Md.): It is a mistake to let AI take over physician duties without final review and approval of a qualified physician. In my experience thus far, physicians, [nurse practitioners and physician associates] are not reviewing and correcting their notes. Is it a time constraint issue or lack of integrity or conscientiousness or lack of medical knowledge?
Many medical records now state at the end of reports and progress notes that AI was implemented to create the notes. All AI-generated notes seem to exhibit similar features: They repeat prior notes, thus adding to the volume of material to be reviewed, three-page notes exceed 36 pages. More significantly, they contain numerous uncorrected errors.
It is uncertain what is causing these errors, whether AI or human. It is obvious that humans are not reviewing and correcting their reports, notes and treatment plans. Common errors include nouns and verbs that are not in accordance with a selected diagnosis/treatment, misdiagnoses on separate occasions for same problem, treatment plan in disagreement with standards of care, treatment plan does not match patient diagnosis/record/profile, an incomplete note missing a treatment plan for a diagnosis, truncated notes, missing relevant items in note (medications, exam, [review of symtoms]), confirmation or change treatment plans missing, duplicated notes one from NP and a second one almost identical signed by MD appearing as duplicate visits/billing for same visit.
As a former medical director for almost 30 physicians, I believe accountability and integrity are of vital importance. AI seems to lead to laziness. If a physician can see more patients by utilizing AI and obtain a bonus, or spend less time on patient care, what good is it if the diagnosis and treatment plan cannot be trusted?
At the Becker’s 32nd Annual Meeting: The Business and Operations of ASCs, taking place October 29-31 in Chicago, ASC leaders, surgeons and healthcare executives will explore strategies to drive growth, enhance operational performance, navigate reimbursement challenges and prepare for the future of ambulatory surgery. Apply for complimentary registration now.