Judgment and Wisdom: Ethical Decision-Making, Evaluating Truth

As machines absorb diagnosis and documentation, value concentrates at the edge of judgment.

The physician's edge moves to where machines are weakest: evaluating truth, weighing evidence, and holding ambiguity and the unprecedented. Judgment here has two faces.

Navigating Complexity

The durable human advantage lies in common diseases with atypical presentations, in contradiction between systems, and in situations outside modeled expectations, where the physician is the escalation point.

The physician is increasingly specializing in the unprecedented: expertise directed at what is novel, ambiguous, or unsolved, engaged where the modeled, the expected, and the stable range give out. AI optimizes the expected. It is weakest where novelty, surprise, or real-time adaptation are required, where no adequate precedent exists, and where even an experienced clinician may not have met the case before. Medicine does not consist only of repeated standard cases. It includes breakdowns, unexpected combinations, edge conditions, evolving states, and trade-offs that are value-laden rather than technical. The role that follows is narrower in one sense and considerably more critical in another. The physician becomes the exception manager, and the human escalation point when advanced systems encounter what they have never seen.

There is a philosophical base. Aristotelian phronesis, practical wisdom, names the situational excellence in judging a particular case that no rule captures, since a rule cannot specify its own conditions of application. Polanyi's tacit knowledge captures it as well: we know more than we can tell, and what cannot be told cannot be transferred into a training corpus. And the distinction between risk and Knightian uncertainty marks the boundary where probability itself gives out. Throughout, the physician, qua human, must remain the final arbiter of truth and of what should be done. The reason is not that machines reason badly. Arbitration is an act that requires someone who can be held to it.

There is yet another formulation of that role: the physician as weigher of evidence and, in the older sense of the word, as judge. The reasoning behind it is a subtraction. If routine care passes to other health professionals, basic diagnostic conclusions to AI, and technical execution to robotics, what remains at the height of the physician’s role is the arbitration of complex questions that depend on more than any of those three can decide. The judicial analogy is exact in the respect that matters. A judge is not the person in the room who knows the most law, but the one who weighs what is presented, decides under conditions where reasonable parties disagree, and is answerable for the decision afterwards. Paired with the meaning-maker role of Chapter VI, this describes the two halves of what the physician does with evidence: weighing what it supports, and rendering what that means for a particular person.

Yet, the analogy should not be overextended. A judge decides between adversarial parties on a record that closes, while the physician decides with the patient on a record that keeps arriving, and may revise the decision next week without anyone having appealed. And a judge is bound to apply the law as it stands to the case as presented, while the physician is free, and sometimes obliged, to reopen what the question even is. The judicial half of the role is real and describes the weighing. The other half is closer to accompaniment, which is why this report calls the physician a guide rather than an arbiter, a distinction taken up below.

Ethical Judgment

Judgment in medicine is never merely technical. Virtue ethics locates moral competence in the settled character of the agent. Principlism, associated with Beauchamp and Childress and dominant in medical education for four decades, works from respect for autonomy, beneficence, non-maleficence, and justice. Casuistry, revived by Jonsen and Toulmin, reasons analogically from paradigm cases. And clinical pragmatism treats moral problem-solving as an experimental, deliberative process. What ties these to the physician is Pellegrino and Thomasma's framing in A Philosophical Basis of Medical Practice (1981), situating ethical reasoning inside the clinical encounter and organizing it around three questions: what can be wrong; what can be done; and what should be done for this patient, which are one act of reasoning.⁵⁹

Machine systems answer the first two questions increasingly well, and the third not at all. That is not because it is harder; it is a different kind of question. What can be wrong and what can be done are questions about the world, answerable from evidence. What should be done for this patient is a question about what is worth wanting, answerable only from a standpoint that has values. The disaggregation of clinical judgment described in Chapter VI runs exactly along this seam. A system that presents an answer to the first two as an answer to the third has made a category error, and only a person positioned to notice will catch it.

Two consequences follow. Ethical judgment becomes a higher-volume activity rather than a rarefied one, because each expansion of the technically possible generates fresh questions about the advisable: germline editing, preemptive intervention on asymptomatic risk, the appropriate use of a simulated presence at a bedside. And it becomes a governance activity as well as a bedside one, since the physician who carries the patient's interest into rooms the patient is not in exercises ethical judgment over model design, escalation thresholds, and procurement, where decisions are made once and affect everyone thereafter.

One such question is nearer to the bedside than the others and is at present largely unanswered. When an opaque system contributes materially to a recommendation, what exactly is a patient consenting to? Informed consent presupposes that the reasoning behind a proposal can be conveyed, at least in outline, to the person who must live with it. A black-box system defeats that presupposition by being unexplainable, whether or not it is wrong. The physician who conveys its output without being able to account for it has technically informed the patient of a conclusion while withholding the thing consent is actually about. The available responses are all imperfect: disclosing the system's involvement and its known error characteristics, declining to rely on systems whose reasoning cannot be inspected at all in decisions above some threshold of consequence, and treating explainability as a medical technology procurement requirement rather than a research aspiration. What is not available is the current default, in which the question is not raised because no one in the encounter is positioned to raise it.

A Word on Placement

Disclosure is necessary and insufficient, since a patient told that an unexplainable system was involved has been informed of a fact rather than of a reason. Explainability as a procurement requirement is the structural remedy and operates too far upstream to help an encounter already underway.

What is needed alongside them is a threshold, and setting it is a governance act of the kind described above. Above some level of consequence, a system whose reasoning cannot be inspected may inform a decision but may not carry it, and the level should be set by the party who must obtain the consent rather than by the party selling the system. Physicians already sit on the procurement and safety committees where such a threshold would be written. Almost nowhere has it been written.

Judgment is where ethical decision-making becomes operational, though it does not belong here alone. It binds together most sectors of the human advantage. It is inseparable from the case for rehumanizing medicine made earlier in this chapter, since medical virtue is what ethical judgment looks like once it has become character.

It runs through complexity and the limits of prediction, because the situations that most demand it are the ones where the model has run out. It underwrites the evaluation of truth, since where a claim came from is a moral question as well as an epistemic one. And it reaches into the consilient and global identities of Chapters VIII and X, into who sits at the table when a protocol is written and whose population supplied the evidence. Ethical judgment cannot be handed to a consultation service and considered handled.

Physicians serve as ethical guides, supplying moral judgment and ethical navigation in care environments that generate more such questions than they resolve. Guide is chosen against two alternatives. An arbiter decides for the patient, which the preceding chapter's account of shared decision-making rules out. A neutral informant declines to help at all, which is an abdication dressed as respect. A guide has walked the terrain, holds a view, and accompanies someone else in reaching their own. The coming decades will require that posture at considerably higher volume than the present ones.

One class of encounter tests that posture more sharply than any other, and it will become more common. Patients accept or reject diagnostic and treatment options on the basis of religious conviction, cultural commitment, personal values, and beliefs the physician may find mistaken. The expansion of what medicine can offer expands the ground on which such refusals occur: predictive testing a person would rather not know the result of; germline decisions affecting descendants; interventions that conflict with a settled view of what a body is for.

The temptation in these encounters runs in two directions and both are failures. One is to treat refusal as a comprehension problem to be solved with better explanation, which mistakes a difference in values for a deficit in understanding. The other is to withdraw into neutrality, which abandons the patient at the moment guidance was actually wanted. The ethical guide establishes what the person is protecting, says plainly what the medicine can and cannot do for it, and remains in the relationship when the answer is no. That last is the hardest part, and it separates a guide from a service provider.

Evaluating Truth

The companion capacity needs more than assertion. The physician's situation is about to become philosophically peculiar: fluent, confident, well-formed claims will arrive in volume from systems with no relation to the world except through a corpus, and patients will arrive having already consulted them.

Philosophy offers no consensus on what truth is, and the disagreement matters in practice. Correspondence theories hold that a claim is true when it matches how things are. Coherence theories hold that it is true when it fits consistently with everything else believed. Pragmatic theories hold that it is true when acting on it works, and clinical reasoning uses all three while rarely distinguishing them.⁶⁰

The distinction matters now because large language models are coherence engines. They generate what fits, fluently and plausibly, with no independent access to how things are. A fabricated citation therefore reads exactly like a real one.

Three further distinctions belong in the physician's working equipment, each routinely violated in practice: between a finding and its interpretation; between absence of evidence and evidence of absence; and between a claim's provenance and its plausibility. Truth is not fluency. The ease and internal consistency with which a claim is expressed carry no evidential weight, though every human heuristic treats them as if they do, and machine output is optimized for exactly those properties.

What the physician must strive for is not certainty, which is unavailable in principle for much of what matters. It is something more modest and more demanding: calibrated, provenance-aware, revisable belief, held by someone willing to be answerable for acting on it.

In practice that means asking where a claim came from and what would have to be so for it to be false; distinguishing the intrinsically unknowable from the merely unknown; and resisting the pull of an answer that is coherent with everything except the patient. This is a discipline of intellectual honesty more than a technique, which returns the argument to virtue: truthfulness is a character trait, and evaluating truth well is what that trait looks like under professional conditions.

A second competence concerns what happens after the physician has reached a view. Discerning real research from fake is the first half of the task. The second is persuading a patient of the distinction, and that is materially harder. Democratized access to data means future patients arrive armed, having done work of their own, sometimes well and sometimes on material that is confidently wrong, and they are evolving in this respect as fast as the physicians treating them.

The skill required is not the assertion of authority, which no longer carries and frequently backfires, but the ability to make provenance legible: to show why one claim is better supported than another in terms the patient can evaluate themselves. Digital bedside manner consists of this, and it can be taught. Alongside it belongs a related capacity the accelerating literature demands: training in how to acquire cross-disciplinary knowledge quickly, which is now as necessary as the ability to appraise it.

The two halves join here. Ethical judgment asks what should be done. Evaluating truth asks what is actually the case. Neither reduces to the other, and both must be exercised by the same person if the decision is to hold together.