An excerpt from Medicine at the Crossroads: Commentaries from a Profession Caught Between Patients, Politics, Technology, and Trust.
In 1979, Kramer vs. Kramer gave us a custody battle in which two wounded adults fought over the future of a child they both loved but understood differently. It was intimate, painful, and unmistakably human. The medical sequel being written today is less likely to unfold in a courtroom than in an electronic health record, an insurance portal, a patient app, or a cloud-based workflow engine. Its title could be Chat v. Chat.
The disputants are not spouses. They are medical chatbots, payer chatbots, pharmacy chatbots, scheduling chatbots, population health chatbots, and patient-facing symptom checkers. Each is armed with its own training data, institutional incentives, workflow rules, and programmed confidence. Somewhere in the middle sits the patient, still nominally the focus of care, but increasingly surrounded by machines negotiating what care is needed, authorized, explained, delayed, or denied.
The idea may sound fanciful, but only if we imagine chatbots as isolated boxes waiting for human prompts. In reality, the more powerful model is autonomous bot-to-bot communication. Non-medical examples are already commonplace: customer service bots escalating to specialist bots, shopping bots bargaining with merchant bots, supply-chain bots interacting with supplier bots, and digital calendar assistants coordinating schedules without human intervention.
Real physician voices, twice a week
Free, and one click to unsubscribe.
A physician recently described a scenario in which a corporate proposal was drafted by one company’s AI, responded to by another company’s AI, and refined through multiple iterations until a finished product emerged. Only after the contract was signed did anyone realize that no human had meaningfully participated in the transaction.
In business, this may be efficient, amusing, or mildly embarrassing. In medicine, it could be fatal. Imagine a patient with worsening shortness of breath. A symptom chatbot collects the history and classifies the problem as low acuity. That summary flows into a scheduling chatbot, which offers the next available routine appointment weeks away. The EHR chatbot reviews the record and finds no red flags because the patient’s language was prematurely parsed into structured fields. Later, a payer chatbot reviews the physician’s order for advanced imaging and denies it because the documented criteria do not satisfy rigid guidelines for medical necessity. The physician appeals, but the appeal is drafted by an internal AI tool using language optimized to satisfy the payer’s algorithm. The payer’s chatbot automatically responds by citing automated policy language. The physician signs, the insurer denies, the patient waits, and the system records each step as a completed success.
Because the system registers this bureaucratic stalemate as a successful administrative resolution, the patient is left completely abandoned in the physical world, still waiting, worsening dyspnea, while their window for early intervention closes. No one in this chain necessarily intended harm. That is precisely the danger. Bot-to-bot medicine may not fail through villainy; it may fail through alignment with institutional routines. One chatbot optimizes for access management. Another optimizes for documentation completeness. Others optimize for denial prevention, cost containment, or patient engagement metrics. Each does its assigned job perfectly. Together, they produce a clinically dangerous outcome that no single person would have designed and no single human can easily explain.
This is where Chat v. Chat becomes more than a clever sequel title. Kramer vs. Kramer was about custody, responsibility, attachment, and the painful question of who was fit to care for a child. The medical version asks who is fit to care for a patient when the custody battle is between competing algorithmic systems. The physician may be told that AI is merely assisting. The payer may claim the final decision remains human. The hospital may argue that clinicians retain authority, and the vendor may insist the tool is not practicing medicine. Yet the practical reality is vastly different. By the time the physician enters the scene, the chatbot conversation may already have shaped the facts, narrowed the options, constructed the medical record, interpreted the policy, and recorded the final decision.
Patients do not present as clean data packets. They hedge, minimize, exaggerate, forget, contradict themselves, and often reveal their most important symptom only after trust develops. Human clinicians learn to hear the pause before the answer, notice the spouse’s worried look, catch the unexpected word choice, and spot the inconsistency that turns a routine complaint into a life-saving diagnosis. A chatbot may capture language but miss context. Worse, one chatbot may pass its incomplete or hallucinated interpretation to another, creating a dangerous illusion of certainty. By the time a physician sees the output, the original patient story has been summarized, categorized, and transformed into fiction with a reduced sense of urgency.
To be sure, some bot-to-bot applications in medicine are entirely beneficial. A pharmacy bot flagging drug interactions within an EHR can prevent injury. A scheduling assistant coordinating post-discharge follow-ups can decrease missed appointments. Chronic disease tools integrating glucose trends, exercise data, and medication reminders can support patients between visits. Similarly, a referral bot transmitting accurate records to a receiving hospital can eliminate important information gaps. However, these automated exchanges must remain strictly accountable to human judgment, clinical context, and professional responsibility.
Preserving that accountability requires five essential safeguards:
- Absolute visibility: Bot-to-bot exchanges that affect diagnosis, triage, treatment, authorization, or follow-up must be fully visible to clinicians and patients. Hidden machine conversations must not dictate care.
- Narrative preservation: Critical clinical decisions must preserve the original patient narrative alongside structured summaries. The patient’s story must not be spread across standardized check-boxes.
- Algorithmic auditing: Systems must document not only the final recommendation but the entire chain of AI-to-AI interactions that produced it.
- Clinical autonomy: Physicians must be in the loop and have meaningful authority to override automated pathways without being penalized by productivity metrics, automated denial workflows, or compliance burdens.
- Mandatory disclosure: Patients have a right to know when autonomous machines are negotiating critical aspects of their medical care.
The future of medical AI should not be a custody battle between bots. A patient is not a project proposal, a claim number, a scheduling problem, or a bundle of structured inputs. The patient is the sole reason the conversation exists. If chatbots are going to square off in medicine, physicians must ensure that the human being in the middle is treated as the focus of care, rather than evidence in someone else’s case.
Arthur Lazarus is a former Doximity Fellow, a member of the editorial board of the American Association for Physician Leadership, and an adjunct professor of psychiatry at the Lewis Katz School of Medicine at Temple University in Philadelphia. He is the author of several books on narrative medicine and the fictional series Real Medicine, Unreal Stories. His latest book is Nobody Told Me There’d Be Days Like These: Hard Truths from Physicians—and What They Mean for Medical Practice.
