Feature Article
By Tom Black, MD
This article was originally published in the September/October 2026 issue of Tarrant County Physician.
A well-known medical maxim says: “When you hear hoofbeats, think of horses, not zebras.” This reminds us that clinical reasoning is based at least partly in probability. After all, the common occurs more commonly than does the uncommon. For generations, this maxim has served as a reasonably reliable guide. But what happens when Artificial Intelligence (AI) hears hoofbeats and suggests that it might be a zebra? And what happens when you ignore that comment?
This is not a hypothetical question. It may instead be the most important medico-legal conundrum of our era. AI has arrived in clinical medicine—not as a future promise, but as an active presence. It is already reading radiographs, flagging indices of sepsis, identifying rare conditions, and pattern-matching across datasets more than a human physician could process in a lifetime. It is the ultimate consultant, opining, advising, and recommending based on the experience of millions of practitioners and cases, not just the few with whom you spoke today in the lunchroom, or even the few dated reports you found on the internet. With the arrival of AI has emerged a paradox so profound that it will undoubtedly reshape the foundations of medical liability. Some legal scholars are calling this the “AI adoption dilemma”: the physician who uses AI and ignores it may be negligent, and the physician who defers to AI may be negligent, but the physician who does not use AI at all may also be negligent. There is, at present, no clearly safe harbor.
To understand the dilemma, we must first appreciate what AI genuinely offers. In every aspect of medical care, error remains a stubborn and costly problem, both in its impact on patient outcome and in cost. Studies have consistently shown that cognitive bias contributes significantly to missed or delayed diagnoses. AI systems are inherently immune to the bias of one’s most recent experience and to fatigue, either of which may cloud one’s judgment. AI does not prejudge on the basis of race, color, national origin, sexual orientation, religion, or age unless instructed to do so. It does not cut corners. It does not assume. In certain specialties such as radiology, dermatology, and pathology, the diagnostic accuracy of AI has in some studies already matched or exceeded that of board-certified specialists. This is not hype. This is peer-reviewed science, and it’s just beginning. For the well-being of our patients, the implications are profound and genuinely auspicious. For the well-being of physicians, the implications are sobering and understandably maybe even alarming.
Here is where legal realities become glaringly apparent. When a physician chooses to use an AI diagnostic tool—whether built into the electronic health record, embedded in an imaging platform, or accessed as a stand-alone clinical decision support system—that tool’s output does not exist in a vacuum. It’s not like reading a medical journal or textbook, asking for a curbside consult, or even consulting online or printed resources. It is timestamped. It is logged. According to ChatGPT, the conversation is a legally discoverable document. If AI notes a finding, and the physician overrides that recommendation without documented clinical justification, a plaintiff has something extraordinarily powerful: an objective, reproducible record showing that the physician was informed and chose not to act. ChatGPT states that the record is kept for 30 days, after which it MAY be deleted, but this does not necessarily occur.
As a result of ongoing litigation involving The New York Times, OpenAI, the owner of ChatGPT, has publicly stated that until reversed, a court order has required it to preserve certain categories of consumer ChatGPT data that otherwise would have been automatically deleted. OpenAI is challenging that order, but until the question is resolved, the usual 30-day deletion schedule may not apply to all consumer accounts. In practical terms, if you do nothing, your conversations may remain stored indefinitely in your account, but even if you actively delete them, they may still be accessible after 30 days. Access to documents would, of course, still require appropriate legal process. Given the current situation, until a ruling has been made, one must assume that anything entered into an online AI service may persist longer than expected due to litigation or regulatory requirements.
The AI situation is categorically different from that involving a human consultant. When a consulting cardiologist suggests additional testing and the clinician overrides those recommendations, there are nuances. A clinical conversation was held, judgments were collegially exchanged, weighing of probabilities took place between two trained minds, and a defensible decision was made. When an AI system notes an anomaly, and the physician dismisses it without documentation, the jury sees something far simpler and far more damaging: an “educated” warning ignored. AI cannot be cross-examined. It cannot explain that the recommendation was tentative. It cannot say it was dealing with a borderline case with nuances. Its recommendation exists unqualified in the medical record. The physician is solely responsible for having overridden the statement.
Now consider the plight of the physician who has simply chosen not to use AI. Perhaps they practice in a manner they consider time-honored: a detailed history, a careful physical examination, and judicious use of evidence-based testing. They have not adopted the AI clinical decision support tool their hospital or employer or university has endorsed. It was not a requirement. No specialty society had mandated it. No state medical board had legislated it as a standard of care. So is the practitioner safe?
The honest answer—for now—is: probably. But that window is closing, and the mechanism by which it will close is one that each of us physicians who is nearing or past the age of retirement has already experienced. I am referring to the adoption of computed tomography. When CT scanning emerged in the late 1970s, it was a remarkable but not universally available technology. A physician in 1980, who chose not to order a CT scan for a patient with new focal neurological deficits, was practicing within the standard of care that still relied heavily on clinical assessment and judgment. By the early 1990s, that same clinical presentation without subsequent CT imaging was becoming increasingly difficult to defend, and by 2000, it was almost universally indefensible. The technology had become so validated, so available, and so clearly superior to clinical assessment alone that its omission in the proper context constituted a clear breach of the standard of care. No law mandated CT scans. No committee voted on the change. The standard had instead shifted organically, driven by wide availability, outcomes data, expert testimony, and the gradual but inexorable consensus of a profession that had witnessed what the tool could do.
AI is on a similar trajectory, but I would argue, with a much steeper adoption curve than that of CT scanning in the 1980s. This is due to the low-cost and universal availability of high-quality AI compared to CT scanners, and to the rapid accumulation of data showing the rapidly increasing efficacy of AI. It is likely that even before the end of this decade, experts in certain specialties will routinely testify that a reasonably competent physician, using the tools available, would and should have employed AI-assisted diagnostic support. At that point, the physician who never utilized AI technology will face a similar or likely even greater degree of exposure as the one who used AI but chose to ignore it.
The situation becomes even more complex when the diagnosis is unclear and may include rare possibilities, outside the probabilities within which humans generally operate. We think we hear a horse because horses are common, but when it’s really a zebra, our mental capabilities may falter. Fortunately for our patients, this is precisely the clinical niche where AI offers its most distinctive value. You read the article in the last edition of this journal by TCMS member David M. Lavine, MD, about his four-year search for an elusive diagnosis, which proved to be cardiac amyloidosis. This is where the legal calculus becomes most pronounced. A physician might reasonably argue that no clinician would have ordered testing for a condition affecting only one in one hundred thousand patients, but if AI suggested that possibility, and the physician dismissed it, that argument collapses. The question becomes not whether a physician would have thought of it, but whether a physician using available technology should have acted on it when AI “thought” of it.
What can the prudent physician do today, in this unsettled and evolving climate? Several practical steps offer meaningful risk reduction, even in the absence of clear legal standards.
Documentation remains the cornerstone of defensible practice, and nowhere is this more crucial than in the context of AI. When a clinical decision support tool generates a recommendation, and the physician chooses not to follow it, that choice must be documented with genuine clinical reasoning. For example, “AI suggested the diagnosis may be a pulmonary embolism, but my clinical gestalt is low given a Wells score of 1, negative D-dimer, and the fact that an alternate diagnosis was confirmed” is defensible. The lack of any documentation explaining your reasoning is not. My mathematically gifted son once failed a math test because he didn’t show every laborious step of his work, having correctly solved each problem in his head. We are to the point where we, too, must show our work or be penalized for not having done so.
Physicians should also understand the regulatory status of the AI tools they use. A tool cleared by the FDA as a medical device is more credible than an experimental algorithm deployed informally. If your institution or employer has an accepted AI platform, not knowing its regulatory standing is unwise.
Staying current with your specialty society recommendations is also essential. No major society has yet mandated the use of AI, but position papers and practice guidelines are beginning to address it. These documents define what that specialty considers to be reasonable practice, and they are precisely the materials that experts will cite in litigation.
Finally, where applicable, physicians should advocate within their institutions for, or demand that their employers enact, policies that are both clinically sound and legally coherent. An AI tool that is mandated without adequate training, support, or clinical validation creates liability without any benefit. Institutions or companies that deploy AI must also bear responsibility, and physicians should insist on clarity about just who is accountable when AI recommendations prove incorrect.
We are at yet another hinge point that will define the practice of medicine for following generations. The AI adoption dilemma—use it and risk liability for not heeding it; use it and risk liability for overreliance; ignore it altogether and risk liability for having not employed it at all—is not a problem that will resolve itself through inaction. It demands engagement from clinicians, medical societies, legislatures, and courts that must eventually provide the clarity that practitioners currently lack.
As a surgeon, I find the scalpel analogy of the title to be illustrative and applicable. The surgeon who picks up a scalpel accepts both its power and its responsibility. The surgeon who refuses to use one when appropriate is not practicing cautious medicine, since not operating may be just as dangerous to the patient—sometimes more so—than performing surgery. AI is becoming, in certain clinical situations, the scalpel of diagnosis. The physician who learns to wield it carefully, and documents its use thoughtfully, ignoring it only with clear clinical reasoning, will be far better positioned—both for their patients and before a jury—than the physician who pretends it does not exist.
Hoofbeats are still unlikely to announce a zebra, but the laws of probability do not work in our favor any longer. AI allows patients to be scrutinized on a much more individual basis than was previously possible, and for a particular patient, the odds of a particular diagnosis, regardless of how common, are either 100 percent or zero. Since AI offers the opportunity for physicians to achieve improved outcomes for our patients, we must embrace it and capitalize on its benefits, but we must also adjust our practices to minimize its risks.
Next: “The Double-Edged Scalpel, Part 2. Precedent Pending”









