The threshold approach rests on a deceptively simple piece of decision theory. For any suspected diagnosis, a clinician can compare the expected benefit of treatment against the expected harm. If the benefit-to-harm ratio is high, treatment is justified even at low probability of disease; if the ratio is low, treatment is warranted only at high probability. These two probabilities define the testing threshold and the treatment threshold. Below the testing threshold, the diagnosis is too improbable to pursue, and further testing adds more risk than information. Above the treatment threshold, the diagnosis is probable enough that the benefits of therapy outweigh the harms of being wrong. Between the two thresholds lies the zone of diagnostic uncertainty, where additional tests are genuinely useful because their results can move the probability across a boundary that changes management.

Rohlfsen and Parsons contend that this elegant structure is being eroded not by new evidence but by the sheer volume of modern clinical information. They invoke entropy in its information-theoretic sense, the average uncertainty contained in a probability distribution. Every laboratory result, imaging study, and vital sign carries information, but the diagnostic value of any single result depends on what the clinician already believes. As the number of available data points multiplies, the average uncertainty in the aggregate picture can rise even when each individual result seems informative. The authors suggest that clinicians and educators have begun to mistake the accumulation of data for the reduction of uncertainty, a confusion that the threshold approach was designed to prevent.

The letter draws on a substantial body of empirical work on diagnostic reasoning. Peter Han, William Klein, and Nikhil Arora, writing in Medical Decision Making in 2011, developed a conceptual taxonomy of uncertainty in health care, distinguishing between the uncertainty of probability, the ambiguity of evidence, and the complexity of the clinical situation. Their framework makes clear that not all uncertainty is statistical. A clinician may know the sensitivity and specificity of a test with precision yet remain uncertain about how those parameters apply to a particular patient whose presentation does not match the populations in which the test was validated. The threshold approach addresses only the probabilistic component, and Rohlfsen and Parsons argue that medical education has overemphasized that component at the expense of the others.

Empirical studies have documented the consequences. Andrew Meyer, Vishal Payne, David Meeks, Rakesh Rao, and Hardeep Singh, in a 2013 vignette study published in JAMA Internal Medicine, examined physicians’ diagnostic accuracy, confidence, and resource requests. They found that diagnostic confidence was often poorly calibrated with respect to actual accuracy, and that physicians frequently requested additional resources even when their stated confidence in a diagnosis was high. The finding is difficult to reconcile with a purely rational threshold model, in which resource requests should scale with the distance between the current probability and the relevant threshold. The authors of the letter suggest that the discrepancy reflects a failure to teach clinicians where their thresholds actually lie, leaving them to set boundaries by intuition rather than by explicit reasoning.

Jeffrey Stojan, Mahan Daniel, Sara Hartley, and Larry Gruppen, writing in Medical Education in 2022, developed a threshold model of dealing with uncertainty in clinical reasoning, examining the roles of experience and task framing. Their work suggests that experienced clinicians do not simply have better-calibrated thresholds; they also frame diagnostic tasks differently, sometimes treating a case as a learning exercise and sometimes as a high-stakes decision, with corresponding shifts in how much uncertainty they are willing to tolerate before acting. The letter argues that this variability is not a defect but a feature of expert judgment, and that pedagogy should make it explicit rather than pretend that a single threshold applies to every situation.

Rohlfsen and Parsons also reference the 5L framework of diagnostic reasoning, developed by Olivia Brumfield, Ahmed Mohamed, Sheng He, and Cory Rohlfsen and published in the Journal of Hospital Medicine in 2026. The framework provides a stepwise scaffold to support clinician educators, breaking diagnostic reasoning into discrete stages that can be taught and assessed. The letter situates the threshold approach within this pedagogical structure, suggesting that the 5L framework gives educators a language for discussing where a diagnostic probability stands relative to a decision threshold, and what kind of evidence would move it. Without such a language, the authors contend, discussions of diagnostic reasoning tend to collapse into vague appeals to clinical gestalt, which students cannot replicate and which resist assessment.

The letter also invokes the ADEPT study protocol, a multicenter prospective quality and safety program described by Jeffrey Schnipper and colleagues in the Journal of Hospital Medicine in 2023, designed to improve diagnostic processes in medical inpatients through prevention and teamwork. The letter suggests that programs like ADEPT represent the systems-level counterpart to the threshold approach, embedding diagnostic reasoning into the structure of hospital care rather than leaving it to the individual clinician. The authors argue that diagnostic excellence requires both the individual judgment of the threshold approach and the organizational support of programs that make diagnostic processes visible, auditable, and improvable.

The implications extend beyond internal medicine. Rohlfsen and Parsons write in the context of predictive medicine and the philosophy of medicine, subjects that the journal’s machine-learning-driven related-articles system associates with the letter. The authors suggest that as diagnostic algorithms become more prevalent, the threshold approach may need to be renegotiated. An algorithm can output a probability, but it cannot set a threshold, because threshold setting depends on the benefit-to-harm ratio of treatment, which is a value judgment about what matters to a particular patient. The letter implies that clinicians who delegate probability estimation to algorithms must retain the judgment required to set thresholds, and that medical education should teach that judgment explicitly rather than assume it emerges from exposure to clinical work.

The letter, published on 8 September 2026, received no funding sources and reports no conflicts of interest. Rohlfsen wrote the initial draft and Parsons reviewed and edited it, with both contributing to the conceptualization of the submission. The work is situated within a tradition of clinical decision theory that stretches back to the Pauker and Kassirer paper, and the authors suggest that the tradition remains vital precisely because the problems it addresses have not been solved by technology. The entropy of modern clinical data has increased, the authors argue, but the fundamental question of when a diagnosis is probable enough to act upon remains a matter of judgment, and judgment is something that must be taught.

Subject of Research: Beyond Entropy: Decision Threshold, Judgment, and Pedagogy

Article Title: Beyond Entropy: Decision Threshold, Judgment, and Pedagogy

Article References: Rohlfsen, C., & Parsons, A. S. (2026). Beyond Entropy: Decision Threshold, Judgment, and Pedagogy. Journal of General Internal Medicine. https://doi.org/10.1007/s11606-026-10753-4

Image Credits: AI Generated

DOI: 10.1007/s11606-026-10753-4

Keywords: Beyond, Entropy, Decision, Threshold, Judgment, Pedagogy, scientific research

Kim, writing from Atlanta at the age of seventy-six and drawing on a formation that spans Korea and America, Christian theology and Chinese thought, describes working with AI as being surrounded by thirty exceptionally capable doctoral students. One offers a term, another restructures the argument, a third suggests what the writer really means, and others return carrying notes from Peirce, Wittgenstein, Confucius, or contemporary sociolinguistics. The experience is exhilarating, and Kim does not deny it. But the metaphor breaks down at a crucial point. Thirty real doctoral students are thirty centers of resistance. One will dislike the argument, another will say Peirce has been misunderstood, a third will accuse the writer of romanticizing East Asia. Someone, sooner or later, will deliver the sentence every scholar needs to hear: the work is beautiful, and it is also wrong.

AI can imitate such objections, but imitation of plurality is not plurality. Current systems, Kim argues, inhabit overlapping textual worlds and share habits of exposition. They are gifted at coherence. Give them two distant ideas and they will often find a bridge before asking whether one ought to exist. This, in the essay’s framing, may be generative AI’s characteristic epistemic danger: the texture of consensus without the evidential weight of consensus. Five machines admiring an idea do not constitute peer review, yet five machines can help that idea put on a suit and tie by afternoon. The old difficulty of scholarship was that a living question might die before a scholar could pursue it. The new difficulty is that a foolish question can become a polished manuscript before dinner.

The essay’s sharpest formulation concerns what AI has actually democratized. Public discussion often claims that these tools have democratized intelligence. Kim counters that they have done something more unsettling: subsidized the appearance of intelligence. Eloquence can now be rented. Erudition can be rented. Structure can be rented. Even the external appearance of judgment can be rented by the hour, or by the token. This does not mean the renter is stupid, and Kim is emphatic on that point, describing a daily practice of using AI gratefully as an extraordinary workshop. A question that once required weeks of preparation can now be tested while it is still warm. Confucian zhengming, the rectification of misalignment between name and reality, can be placed beside Peircean semiotics, Wittgenstein can be invited to disturb the arrangement, and a sociolinguist’s objection can be summoned in seconds. The library is illuminated at three in the morning. The translator is awake.

Speed, Kim acknowledges, preserves the temperature of a question, and that is no small gift. But speed also creates a new temptation: linguistic completion before intellectual completion. The machine closes sentences, fills conceptual gaps, smooths awkward transitions, and reconciles tensions with extraordinary politeness. What once remained visibly unresolved can now be made to look settled. Fluent closure has become almost free, and the institutional consequences follow directly. When half-formed thoughts can be dressed for submission in an afternoon, journals inherit part of the cost that AI has removed from authors. Editors and peer reviewers become unpaid inspectors of premature coherence. The manuscript becomes cheaper to produce precisely when judgment becomes more expensive to supply.

Kim’s personal response is to cultivate a virtue that sounds almost trivial but carries considerable technical weight: the ability to say almost. Almost right. Almost clear. Almost mine. Almost worthy of publication. The discipline involves a battery of resistant responses to fluent output: the draft is too smooth, the problem has been Westernized, the concepts should not be reconciled yet, the word is technically correct but existentially wrong, the English has been improved while the thought has been weakened, the concept is right but the center of gravity is wrong. Kim increasingly suspects that this capacity will matter more than brilliant prompting. Prompt engineering, in the essay’s wry framing, is becoming a minor priesthood, complete with incantations and claims to esoteric skill. Learn the spells by all means, but the deeper human competence is recognizing when an elegant answer has betrayed the question.

The problem is especially acute for thinkers who work across languages and civilizations. A thought may arrive with a Korean emotional temperature, a Chinese conceptual skeleton, a Christian theological horizon, and a Western philosophical question, and academic English asks it to pass through customs. AI is a magnificent customs broker, Kim writes. It can move the cargo quickly. It can also quietly replace the cargo. To guard against that substitution, Kim has adopted a deliberately primitive discipline: before inviting AI into an important thought, two or three ugly sentences of the original intuition are preserved on the page, sometimes in Korean, sometimes grammatically incomplete. The purpose is not legal ownership, since the mythology of the solitary author was always exaggerated and people think with teachers, enemies, books, languages, lovers, traditions, and ghosts. The purpose is intellectual provenance, the ability to trace a refined phrase back to the crude perception from which it grew.

That provenance question is illustrated with a phrase from Kim’s own writing, indexicality without occupancy, and the Korean intuition behind it: AI speaks like a person of authority, but there is no one behind the words to bear responsibility; there is a sign, but no person there. The important question, Kim argues, is not whether the machine offered the elegant phrase or whether the human refined it through forty exchanges, but whether the language clarified the original intuition or replaced it. Formal governance will not settle this. Academia will need disclosure rules for AI use, publishers will invent policies, and universities will produce forms, all necessary and all predictably bureaucratic. The more intimate problem will remain unregulated: whether scholars will still recognize their own thoughts after the machines have improved them.

Kim is less worried that AI will become intelligent than that human beings will become satisfied with the signs of intelligence, and academia already rewards those signs: a polished abstract, correct vocabulary, a symmetrical framework, six boxes in a diagram, a confident conclusion beginning with a call for fundamental reconceptualization. AI did not invent academic vanity, the essay observes; it simply industrialized some of its preferred costumes. The antidote is not nostalgia, and Kim explicitly refuses to dismiss a technology that enables what the essay calls speculative adjacency, bringing distant bodies of thought into rapid, exploratory contact. When the hallways become short enough, the architecture of thinking changes. But the cheaper inquiry becomes, the more expensive discernment must become. Not every connection deserves development, not every curiosity deserves a paper, and not every publishable paragraph contains a thought worth publishing. In the old economy of scholarship, scarcity imposed crude discipline through limited books, research trips, conversations, and years. AI has removed many of those restraints, and what is needed now, Kim writes, is intellectual fasting.

The essay closes on a question that reframes the entire debate. Rather than asking whether AI will make us more intelligent, a question that flatters machine and human alike, Kim asks whether, surrounded by nearly free fluency, we are becoming more discriminating. Are our questions improving? Can we recognize a false harmony? Can we preserve an inelegant intuition long enough to discover what it means? Can we throw a curveball that breaks the window of premature coherence and then inspect the glass? Above all, can we still say almost when the machine gives us exactly the beautiful answer we hoped to hear? That small word, Kim suggests, may become one of the last defenses of formed intellect. AI has made fluency cheap. Judgment is now expensive, and we should spend accordingly.

Subject of Research: The epistemic impact of generative AI on human judgment, intellectual provenance, and scholarly writing

Article Title: The ability to say “almost”

Article References: Kim, S. (2026). The ability to say “almost”. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03297-y

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03297-y

Keywords: generative AI, artificial intelligence, epistemology, academic writing, peer review, judgment, prompt engineering, intellectual provenance, AI & Society, scholarship, language models, critical thinking