This article emerges from a two-day workshop at UiT The Arctic University of Norway organized by Alexandra Brandt Ryborg Jønsson in May 2026. Key panelists were Alexandra Brandt Ryborg Jønsson, Stefan Ecks, Dorthe Kristensen, Stine Willum Adrian, and John Brandt Brodersen. The empirical examples discussed here are drawn from the workshop participants’ presentations: Jønsson on ageing, chronic illness, and loneliness; Kristensen on AI in radiology; Willum Adrian on prenatal screenings for congenital heart defects; Brodersen on overdiagnosis; and my own ethnography of how mental illness is diagnosed in India.
Abstract
Every major theory of medicalization and diagnostic harm has proceeded from a shared assumption: that the problem starts when a classification enters a patient's life. This article argues that such theories have missed looking at what happens before that. Between the observable signal and the patient encounter lies a threshold that has remained conceptually invisible: the intra-biomedical decision about whether a specific clinical finding (a clot, a node, a shadow) should become a biomedical classification towards action. Drawing on Living Value Theory (LVT), the article distinguishes two analytically separate thresholds. The first is internal to biomedicine: should a clinical signal become a diagnostic category? The second concerns descent: should that biomedical category enter the patient's mesocosm? These are distinct decisions with distinct logics, distinct actors, distinct ethical stakes, and distinct harm profiles. The article develops this two-threshold architecture through four applications: a taxonomy of diseases by recursive structure; the case of psychiatry as the domain where classificatory objects move during stabilization itself; the introduction of AI into diagnostic radiology as a new recursive agent operating primarily at the first threshold; and the ethical geometry that results from properly distinguishing the two thresholds. The conclusion proposes mesocosmic literacy as the methodological contribution LVT makes to overdiagnosis research, clinical practice, and health policy.
Keywords: overdiagnosis, medicalization, Living Value Theory, recursivity, diagnostic iatrogenesis, AI in medicine, mesocosm, psychiatric classification
I. The Invisible Threshold
Consider a radiologist in a reading room looking at a scan. No patient is present. No patient has complained of anything. No felt misalignment has moved upward through any living body to request clinical attention. There is only an image on a screen and a professional looking at it. What the radiologist sees -- a shadow, a density variation, a subtle asymmetry, an irregularity that might be nothing at all -- exists, in Living Value Theory's terms, at L3: a symbolic articulation of something registered by a detection apparatus. What the radiologist must now decide is whether that L3 observation should become an L4 biomedical stabilization. Should this be reported? And if reported, as what? A finding requiring follow-up? A risk marker warranting surveillance? A diagnosis requiring immediate intervention? Or a normal variation, a coincidental finding, a noise artifact, something that should not enter any clinical record at all?
This decision -- signal or noise, variation or pathology, risk or irrelevance, monitor or ignore -- is made before any patient encounter. It is made entirely within the biomedical apparatus. Its consequences for any particular human life are already potentially enormous, and yet no patient's body has produced any signal requesting it. The disruption, if there is to be one, will originate here: in this reading room, in this professional judgment, before a single clinical conversation has occurred.
This is the first threshold, and it is the one that theories of medicalization, overdiagnosis, and diagnostic harm have most consistently overlooked. The medicalization critique beginning with Illich and developed through Conrad, Zola, and their successors has always begun from the clinical encounter: the expansion of medical authority into domains of ordinary life, the extension of diagnostic categories beyond their appropriate scope, the colonization of suffering by professional classification. The overdiagnosis literature in clinical medicine and epidemiology has similarly focused on populations of diagnosed patients, on rates of detection in excess of rates of genuine disease, on harm produced by treatment of conditions that would never have caused symptoms. Both traditions have identified real and consequential problems. But both begin their analysis at the second threshold: the moment at which a biomedical category enters a patient's world. I would argue that the first threshold, as a distinct site of decision, analysis, and potential harm, has not been given enough attention yet.
I propose that this is a structural gap rather than an oversight. Every existing framework begins too late because every existing framework begins from what announces itself -- from suffering already present, from illness already felt, from classification already installed. The space before the first threshold, the moment of intra-biomedical decision in which a clinical observation either becomes or fails to become a diagnostic category, has remained outside the field of theoretical attention. Naming it and developing its architecture is the primary task of this article.
The stakes of that task are not merely theoretical. They have become urgent at a specific historical moment, because the first threshold is currently being transformed by a new kind of agent: artificial intelligence. Diagnostic AI systems in radiology, pathology, cardiology, and other fields operate precisely at the first threshold. They are designed to identify, flag, and classify L3 signals before any clinician has seen them and before any patient has been informed. The introduction of these systems raises questions that no existing framework is equipped to answer, because no existing framework has adequately theorized the threshold at which they intervene. The consequence, as will be argued in section VI, is that AI diagnostic systems are being evaluated almost exclusively through accuracy metrics at the moment when the far more consequential question concerns their recursive effects on clinical judgment, institutional thresholds, and eventual mesocosmic descent.
The argument proceeds as follows. Section II provides a condensed account of Living Value Theory's architecture, which the remainder of the article presupposes. Section III develops the full two-threshold model and demonstrates that the two thresholds are analytically distinct. Section IV proposes a taxonomy of diseases and conditions by recursive structure. Section V develops the case of psychiatry as the domain in which the recursive instability of classificatory objects is most acute. Section VI analyzes AI diagnostic systems as new recursive agents operating at the first threshold. Section VII develops the ethical geometry that results from properly distinguishing the two thresholds. Section VIII proposes mesocosmic ontological literacy as the methodological contribution LVT makes to the field.
II. Living Value Theory: A Condensed Account
Living Value Theory describes the architecture of living coordination -- how embodied beings engage with their world, how disruptions to that engagement become available to themselves and to institutions, and how institutional responses affect the coordination they claim to address. A full exposition is available elsewhere (Ecks 2022, 2026; livingvaluetheory.org). What follows is the condensed account required for the argument of this article.
LVT proposes five levels of recursivity through which coordination becomes available to itself, to others, and to institutions. Level 1 (L1) is seamless coordination: the primary level at which life proceeds without friction, deliberation, or symbolic articulation. The body compensates and adapts. Familiar environments afford movement without thought. Social rhythms proceed without renegotiation. Health at L1 is most fully operative when it leaves the least trace -- which means it is, within any system organized to detect health through events, a permanent structural absence. Not because it is not there, but because it is nowhere that such a system can look.
Level 2 (L2) is felt misalignment: the pre-symbolic sense that something is off. Not yet named, not yet narrated, simply felt. Most L2 states are remediated immediately through spontaneous bodily self-correction: a shift in posture, a pause, a reach for water, a reduction of pace. This continuous felt self-management -- L2 remediation without symbolization -- is health management at its most primary. It requires no clinical encounter, no institutional authorization, and it leaves no transactional trace.
Level 3 (L3) is symbolic articulation: something becomes nameable, communicable, expressible. "I have pain here." "Something does not feel right." "The scan shows something unusual." L3 is where the transition from pre-symbolic to symbolic occurs. Level 4 (L4) is institutional stabilization: symbolic articulations are organized into general categories, diagnostic classifications, decision-guiding thresholds, and population-level protocols. "You have major depressive disorder." "Your bone density falls within the osteopenic range." "There appears to be a small aneurysm." These are not simply descriptions. They are actionable stabilizations that exist precisely to drive decisions. Level 5 (L5) is meta-recursive reflection: the entire system of levels becomes an object of analysis. This article operates at L5.
Coordination occurs through five irreducible mediations: embodiment, being-with, dwelling, multimateriality, and multisymbolization. None can be derived from another. Any intervention that enters a mesocosm does not touch one mediation -- it reorganizes all five simultaneously. Embodiment is the lived body as the ongoing condition of all experience: sensation, proprioception, metabolism, fatigue, pain. Being-with is coordination with other living beings: co-presence, attunement, dependency, care. Dwelling names the spatial and temporal conditions of inhabitation: landscape, climate, seasons, light cycles. Multimateriality names coordination through materials and artefacts: tools, medicines, infrastructures, monitoring devices. Multisymbolization names coordination through symbols -- language, number, classification, law -- and is the mediation most prone to overreach: the systematic mistake of treating symbolic representation as if it were the coordination it represents.
The directional principle is the hinge on which the argument of this article turns. In healthy living, recursivity flows upward: L1 coordination sustains itself; disruption is felt at L2; if it persists it is named at L3; if naming is insufficient, L4 institutional categories are brought in. The institution responds to what the body has already signalled. In preventive diagnostics, the direction is inverted: L4 classifications descend onto L1 in the absence of any prior L2 signal. The institution arrives first, installs its classification, and that classification reorganizes L1 from above. "What Medicine Cannot Register" (Ecks 2026) developed this directional principle and named the harm it generates as preventive diagnostic iatrogenesis: harm generated by the downward installation of L4 classifications onto L1 coordination that has produced no L2 ground for them. That article treated the downward movement as a single event. The present article argues that it comprises two distinct events -- two thresholds, not one -- and that the distinction between them has consequences for theory, methodology, clinical ethics, and health policy that the single-event account cannot produce.
III. The Architecture of Two Thresholds
The first threshold is intra-biomedical. It concerns the decision about whether an L3 clinical observation should become an L4 biomedical stabilization. This decision is made entirely within the biomedical apparatus, by biomedical actors, using biomedical criteria. The patient is not yet involved. The patient's mesocosm has not yet been entered. The question at this threshold is: should this signal be stabilized into a category?
The questions that constitute this threshold include: Is this noise or signal? Is this variation within normal limits or a departure from them? Is this a risk that requires monitoring, or an anomaly requiring immediate action, or a finding that should not be recorded at all? Is this an artifact of the detection technology or a feature of the body being examined? Does the probability that this finding represents genuine pathology exceed the threshold at which clinical response is warranted? These are not simple questions. They require judgment about sensitivity and specificity, about base rates in the population, about the natural history of conditions that resemble this finding, about what follow-up investigations would cost in terms of resources, patient burden, and cascade effects. They are questions that radiologists, pathologists, cardiologists, and clinicians in many specialties answer daily, often under time pressure, often with incomplete information, and almost always without explicit frameworks for weighing the recursive consequences of their decisions.
What the first threshold has not had is a theoretical framework that identifies it as a distinct site of analysis. The overdiagnosis literature has shown that certain conditions are systematically over-detected -- that population screening programmes find pathology at rates that exceed rates of symptomatic disease, that many detected conditions would never have caused harm if left undetected, and that treatment of screen-detected conditions produces net harm in identifiable populations. These are vital findings, but they focus on existing data, like rates of detected disease and rates of downstream harm. They have not yet focused on how the first threshold is produced.
The second threshold concerns descent: should a biomedical L4 category enter the patient's mesocosm? This threshold is not clinical in the narrow sense -- it is not simply the moment of clinical communication. It is the entire set of decisions and processes by which a biomedical stabilization becomes a mesocosmic event: is the patient told? In what terms? With what framing? With what information about natural history, prognosis, and the limits of clinical knowledge? And if told, what then? Is active treatment recommended? Surveillance? Watchful waiting? Lifestyle modification? Nothing?
The descent, if it occurs, reorganizes the patient's mesocosm across all five mediations simultaneously. Embodiment: the body that had been coordinating without demanding attention becomes a body under permanent self-surveillance. Being-with: relationships are reorganized around the disclosure, or around the secret of a finding the patient does not share. Dwelling: inhabited spaces are inhabited differently under the anticipatory weight of a possible future. Multimateriality: a diagnostic or surveillance programme installs a material infrastructure -- clinic appointments, monitoring devices, medications, imaging equipment -- that was not generated by anything in the patient's prior material coordination. Multisymbolization: a word, a category, a diagnosis, reorganizes the patient's symbolic self-understanding permanently and without guarantee of return.
Leif, documented in ethnographic fieldwork among older Danes (Jonsson 2024) and analyzed in detail in "What Medicine Cannot Register" (Ecks 2026), illustrates the second threshold with precision. At 58, he underwent a routine PSA test. Nothing in his lived coordination had generated a signal requesting it. The test found prostate cancer. He was enrolled in a surveillance programme that followed him for fifteen years. The cancer never grew. He died of a stroke. The surveillance programme, within the biomedical apparatus, was a success: it monitored a condition and produced no unnecessary treatment. Traced across all five mediations, the harm it produced was enormous: fifteen years of altered embodiment, maintained social secrecy, anticipatory reorganization of dwelling, a material surveillance apparatus installed in daily life, and symbolic self-understanding permanently oriented toward anticipated death. None of this registers as a clinical finding. The disruption of a life that was, before the PSA test, simply working is constitutively illegible within the framework that caused it.
But now notice what Leif's case requires. The descent into his mesocosm -- the second threshold -- was consequential on a prior decision: the decision to communicate the finding, to enrol him in surveillance, to make the prostate cancer an active part of his clinical and personal life. That decision was itself consequential on a prior decision: the decision to perform the PSA test in the first place, and then to classify the finding as a condition warranting clinical management rather than a variation to be noted and disregarded. That prior decision is the first threshold. And it was made, entirely within the biomedical apparatus, before Leif was ever told anything.
The two thresholds are therefore not merely analytically distinct. They are temporally and institutionally separate. Different actors make them. Different criteria govern them. Different bodies of evidence bear on them. Different ethical frameworks apply to them. And crucially, the harm profiles they generate are different in kind, not merely in degree.
At the first threshold, the harm is the harm of generating an L4 stabilization that should not have been made: a signal classified as a condition when it should have been classified as noise, variation, or a finding below the threshold of clinical significance. This harm is invisible at the moment of decision because it has not yet touched any patient's mesocosm. It becomes visible retrospectively, through outcome studies, through overdiagnosis research, through the accumulated evidence that a given screening programme produces more harm than benefit. But it cannot be seen in the individual case at the moment of decision.
At the second threshold, the harm is the harm of descent: the reorganization of a mesocosm by a classification that had no L2 ground. This harm is also invisible within the biomedical apparatus, because it is distributed across mediations that the apparatus does not track and because it produces no discrete event that a record system is built to capture. But it is felt in the mesocosm. It reorganizes daily life. It changes who a person is.
The two harms are not additive. An error at the first threshold makes an error at the second more likely. A correct decision at the first threshold does not guarantee a correct decision at the second. And a careful, well-informed decision at the second threshold cannot undo harm generated by a poorly made decision at the first. The thresholds are connected but not equivalent, and treating them as a single event -- as all existing frameworks do -- produces systematic theoretical, empirical, and ethical errors.
The directional principle, restated in light of this architecture, now reads as follows. In genuine illness, disruption flows upward: L1 coordination is disrupted, L2 misalignment is felt, L3 articulation follows, and L4 institutional response is requested from below. The entire sequence is initiated by something that has already moved in the mesocosm. In preventive medicine, there are two downward movements rather than one. The first movement is intra-biomedical: an L3 signal is stabilized into an L4 category by professional judgment within the biomedical apparatus. The second movement is mesocosmic: that L4 category descends into a living person's world. Each movement is distinct. Each is contingent. Neither is entailed by what precedes it. And the harm at each site is different in character from the harm at the other.
IV. Disease Taxonomy by Recursive Structure
Once the two-threshold architecture is established, it becomes possible to develop a taxonomy of conditions by recursive structure that has not previously existed. Existing disease taxonomies are organized by organ system, aetiology, pathophysiology, or clinical presentation. None of these dimensions captures what is, from the perspective of the two-threshold model, the most consequential variable: the recursive profile of the condition, meaning how much recursive instability it generates intrinsically, and how much it generates through descent into the mesocosm.
Two primary types can be distinguished, with important variations within each.
The first type consists of conditions that are biologically minimally recursive but mesocosomically massively explosive once communicated. Small incidentalomas, tiny aneurysms, indolent prostate cancers, low-risk genetic variants, ductal carcinoma in situ, subclinical hypothyroidism, indeterminate pulmonary nodules -- these are conditions in which the biological process, left undetected, proceeds or does not proceed without generating significant L2 disruption. The underlying biology is not recursively unstable: it does not change in response to being known about. An aneurysm does not enlarge differently because the person who carries it has been told it is there. A prostate cancer classified as low-grade does not accelerate because a diagnosis has been communicated. The biological object is, in the relevant sense, non-recursive with respect to its classification.
But once the detection is made, once a category is installed, the mesocosmic consequences are potentially enormous and are always recursive. The word "cancer" does not sit quietly alongside a person's existing sense of themselves. It reorganizes that sense. Anticipatory anxiety, altered bodily awareness, changed relationship to time, reorganized social bonds, new material surveillance apparatus: all of these are recursive consequences generated not by the biological process itself but by the descent. For conditions of this type, the harm profile is almost entirely a second-threshold phenomenon. The question at the first threshold -- should this observation be stabilized into a category? -- is therefore also a question about whether the mesocosmic consequences of descent are worth the clinical benefits of early detection. That calculation cannot be made without understanding what descent actually does. And it cannot be made from within a framework that has not distinguished the two thresholds.
The second type consists of conditions that are intrinsically recursively unstable: conditions in which the classification itself partly reorganizes the phenomenon it claims to name. Depression, ADHD, loneliness, burnout, post-traumatic stress disorder, premenstrual dysphoric disorder, anxiety disorders -- these are conditions in which the act of diagnosis is itself an event that changes what is being diagnosed. The diagnosis reorganizes the patient's self-understanding, which reorganizes their bodily experience, which reorganizes what they report to clinicians, which reorganizes subsequent diagnoses. The recursive loop is not a side effect of classification. It is a constitutive feature of the domain.
For conditions of this type, the first threshold is not simply a decision about whether to record a signal. It is a decision about whether to initiate a process of recursive self-reorganization in a person whose current coordination may not require it. The L3 signal that precedes the first-threshold decision may itself be recursively generated: someone presenting to a clinic with fatigue, low mood, and difficulty concentrating may be presenting from a life situation in which these states are entirely appropriate responses to genuine problems, rather than symptoms of an underlying disorder. The first-threshold decision to stabilize this presentation into a diagnostic category sets in motion a chain of events that will change what the person experiences, how they describe it, and what clinical responses it subsequently attracts.
These two recursive types -- biologically static but socially explosive, versus intrinsically recursively unstable -- have very different implications for both thresholds. At the first threshold, the question for biologically static conditions concerns the probability of genuine pathology and the expected natural history. For intrinsically recursive conditions, the question is different in kind: does this presentation constitute a state that genuinely requires institutional stabilization, or is it a response to life circumstances that would reorganize itself if those circumstances changed? At the second threshold, the question for biologically static conditions concerns the expected recursive consequences of descent given what is known about natural history. For intrinsically recursive conditions, descent is itself part of the natural history, because the classification changes what the condition subsequently becomes.
A further distinction within the second type deserves attention. Some intrinsically recursive conditions -- depression is the paradigm case -- become partially less recursive as biological correlates are identified and effective pharmacological treatments are developed. The recursive instability does not disappear, but it is partially constrained by treatment response. Other conditions -- loneliness is the clearest current case -- are epidemiologically measured and clinically categorized in ways that generate recursive consequences while offering no corresponding clinical benefit. The epidemiological finding that lonely people have elevated mortality risk, reported in popular media as a health imperative, produces the recursive descent of a guilt-laden self-assessment into ordinary social life. The person eating dinner alone now inhabits their dinner as a health risk. The categorization generates the recursive harm of descent without any of the compensating benefits that, in other domains, might justify it. The L4 epidemiological category has descended into L1 social life without passing through any first threshold at which its clinical significance was evaluated.
This taxonomy by recursive structure is not merely of theoretical interest. It has direct implications for the design of screening programmes, for clinical decision-making, and for the ethics of population health surveillance. Screening programmes designed for biologically static conditions need to address second-threshold consequences explicitly, because harm is generated almost exclusively there. Programmes extending L4 classifications to intrinsically recursive conditions need to address whether classification produces net benefit given its constitutive role in reorganizing the phenomena it names. The taxonomy makes visible a set of distinctions that existing frameworks have no vocabulary to capture, and that are therefore not being systematically applied in clinical or public health practice.
V. Psychiatry as Limiting Case
Psychiatry occupies a structurally distinct position in the two-threshold model because its classificatory objects are intrinsically recursively unstable to a degree that no other medical domain approaches. Understanding why this is so is important not only for psychiatry but for what psychiatry reveals about the limits of biomedical classification more generally.
In most biomedical domains, the relationship between the classified object and the classification is asymmetric in a specific way: the classification does not change the object. An aneurysm does not alter its biological course in response to being labelled an aneurysm. A tumour does not grow or regress differently because of what it is called. A kidney stone does not move because the diagnosis has been communicated. The object is, in the relevant sense, non-recursive with respect to its own classification. This is what makes quantification tractable in these domains: the measuring instrument does not interfere with what it measures, or interferes only minimally and in quantifiable ways.
Depression does not behave like this. A diagnosis of depression is itself an event that changes the course of whatever is being diagnosed. The person who receives the diagnosis reorganizes their self-understanding around it: they now have a condition with a name, a prognosis, a treatment pathway, a community of fellow sufferers, an established pharmacological response, and a set of expectations about trajectory. All of these reorganize what the person experiences, how they report it, what they seek from clinical encounters, and how they respond to treatment. The classification has entered the phenomenon and is changing it from within.
This is why every serious puzzle in psychiatric pharmacology becomes intelligible from this perspective. The large and variable placebo effects in antidepressant trials are not measurement noise: they are the recursive consequences of being in a context of therapeutic attention, organized expectation, and symbolic confirmation of suffering. The declining efficacy of apparently unchanged drugs across decades -- a phenomenon that has generated substantial methodological controversy -- is not statistical artifact: it is the recursive consequence of changing diagnostic cultures, changing patient expectations, and changing clinical practices, all of which alter the condition being treated even as the molecule remains constant. The poor translation of clinical trial results to real-world practice -- drugs that perform well in trials performing far worse in ordinary clinical use -- is not a problem of patient selection or adherence: it is a consequence of the fact that the trial context is itself a recursive intervention that changes what the drug is treating.
None of these puzzles can be resolved by better trial design. They are not methodological problems that improved methods could in principle overcome. They are structural features of a domain in which the objects of study are recursively unstable with respect to the investigations being conducted on them. Once a phenomenon can be fully quantified, it is, in the relevant sense, no longer fully recursive. And psychiatric conditions, insofar as they are constituted through self-understanding, social meaning, and symbolic self-organization, will always resist full quantification in principle, not merely in practice. This is why the most important things in psychiatry -- the mesocosmic coordination of actual lives -- will never show up cleanly in a clinical trial.
This has a specific implication for the first threshold in psychiatry. The intra-biomedical decision about whether an L3 clinical observation should become an L4 psychiatric stabilization is not analogous to the corresponding decision in radiology or oncology. In radiology, the first-threshold decision concerns whether a biological finding meets a clinical significance threshold. The finding exists independently of the decision; the decision does not change the finding. In psychiatry, the first-threshold decision is constitutively involved in the phenomenon it purports merely to classify. The decision to stabilize a presentation as major depressive disorder rather than as a normal response to loss, as ADHD rather than as a developmentally appropriate pattern of attention in a particular child, as personality disorder rather than as a coherent self-protective response to an adverse social environment -- these decisions are not simply readings of pre-existing states. They are interventions in recursive processes that are partly constituted by what they are called.
The clinical portraits from fieldwork in Kolkata, documented in Living Worth (Ecks 2022), illustrate this with a precision that theoretical exposition alone cannot achieve. Dr. Mullick, the high-flying psychiatrist who travelled several hours daily across the city to practice in different clinics, made diagnoses that were correct by all biomedical criteria. He stabilized genuine L3 presentations into appropriate L4 categories. He prescribed indicated medications. And yet the consequences in the cases he managed -- the ADHD diagnosis that organized a family's violence against a mother, the psychiatric labels that drove a woman to sleep under a stairway -- were catastrophic. The first-threshold decisions were clinically defensible. The second-threshold consequences were, in the actual mesocosms into which those decisions descended, devastating.
Mrs. Saha's case is worth tracing precisely. Her son Neel, then twelve, had been struggling at school. Dr. Mullick diagnosed ADHD and prescribed methylphenidate. The diagnosis entered the household like a weapon. Mrs. Saha's husband and in-laws decided that if anyone in the family was mad, it was Mrs. Saha. They turned on her. Four years after Neel's diagnosis, Dr. Mullick diagnosed Mrs. Saha herself with manic depressive psychosis. Her husband and his brother began to beat her. She was exiled to sleep alone under the stairway. By the time she was observed in fieldwork, she had one goal: to obtain a brain scan that might prove her condition was neurological rather than psychiatric, and thereby restore some standing in the family. Dr. Mullick, when he described the entire trajectory -- the ADHD diagnosis, the family's reaction, the beatings, the exile -- showed no recognition that any of this had anything to do with what he was doing. He was practicing by the book. He had made correct diagnoses. He had prescribed appropriate drugs. He knew the facts. He did not know the mesocosm.
Dr. Sen, by contrast, practiced from a small chamber at the edge of a slum, a short walk from the outpatient hospital where Dr. Mullick operated. He had known many of his patients for years. His decision not to prescribe antidepressants to a severely depressed woman whose husband was an unemployed alcoholic and who had to maintain her whole family on her own was not a violation of clinical guidelines. It was a recognition that the second threshold could not safely be crossed in this particular mesocosm: that the medication would reduce the patient's capacity to work, strain the household budget, and risk destabilizing the fragile coordination holding the family together. Pills, he said, would bring down the house. The most valuable clinical judgment he exercised left no trace in any record. It was, from the perspective of the two-threshold model, a correct decision at the second threshold that depended on mesocosmic knowledge that no protocol could supply.
The comparison between Dr. Mullick and Dr. Sen is not a moral one. Dr. Mullick is a good psychiatrist by all institutional standards. The point is ontological: good clinical judgment at the second threshold requires mesocosmic knowledge, and mesocosmic knowledge requires mesocosmic exposure -- actual, metabolic, embodied presence in the world of the patient. This is not something that can be reduced to cultural competency training or additional modules in the medical curriculum. It requires a different understanding of what clinical knowledge actually is, and where it lives.
VI. Artificial Intelligence as New Recursive Agent at the First Threshold
The introduction of AI diagnostic systems into clinical medicine represents a qualitative transformation of the first threshold, not merely a quantitative improvement in its execution. Understanding why this is so requires analysis that existing frameworks cannot provide, because those frameworks have not theorized the first threshold as a distinct site of analysis.
Current evaluations of AI diagnostic systems proceed almost exclusively through accuracy metrics: sensitivity, specificity, positive predictive value, negative predictive value, area under the receiver operating characteristic curve. These metrics answer the question of how well a system identifies true positives and true negatives in a defined population with respect to a defined reference standard. The question is not unreasonable. It addresses a genuine concern about whether AI systems can perform comparably to or better than human clinicians at pattern recognition tasks. And in some settings, the evidence that AI systems can match or exceed human performance at these tasks is genuine and important.
But accuracy metrics answer only a narrow version of the question that matters. They measure the system's performance at the first threshold under conditions in which the correct answer has already been determined through some reference standard. They do not address what the system does to the first threshold itself -- how it changes the conditions under which clinical judgments about L3-to-L4 stabilization are made. This is where the genuinely consequential effects reside, and it is a question that existing evaluation frameworks have no vocabulary to address.
AI diagnostic systems operating at the first threshold introduce at least four distinct recursive transformations. First, they generate new L3 signals. An AI system trained to detect subtle features of radiology images may identify findings that human radiologists would not have noticed, or would have noticed and dismissed as below the threshold of significance. These AI-generated findings exist at L3: they are symbolic articulations of something the system has registered. But they carry a different epistemic character from human-generated L3 findings, because their generation is opaque, their reliability for findings at the margins of sensitivity is uncertain, and the clinical significance of findings that only AI systems can reliably detect is often unknown. The first-threshold question -- should this L3 observation become an L4 stabilization? -- is therefore being asked about a new population of signals whose clinical significance has not been established by any route other than AI detection itself.
Second, AI systems change the conditions under which human clinicians exercise first-threshold judgment. The radiologist who knows that an AI system has flagged a region of an image is no longer making the same judgment as a radiologist examining the image without that information. The flag changes what is attended to, how it is interpreted, and what threshold of certainty is required before recording or dismissing a finding. This transformation of clinical judgment by AI is not simply a matter of adding information. It reorganizes the epistemic situation in which judgment is exercised. It may raise effective sensitivity thresholds, making radiologists more willing to record borderline findings that the AI has also flagged. It may shift responsibility, making clinicians reluctant to dismiss findings that the AI has identified, because the institutional and medicolegal consequences of a false negative following AI flagging are different from the consequences of a false negative made without AI input.
Third, AI systems at the first threshold create cascade risks that are structurally different from those generated by human-only first-threshold decisions. If an AI system is deployed across a large screening programme and achieves a false positive rate of even one or two percent, the absolute number of people for whom an L4 stabilization is created without genuine L3 ground may be enormous. Each such stabilization creates the conditions for second-threshold descent. The question is not whether any individual AI-generated finding is clinically significant, but what the population-level consequences of the recursive cascade are when AI generates first-threshold decisions at scale. These consequences cannot be assessed in advance, because they depend on how clinicians respond to AI flagging, how institutions adapt their protocols, how patients interpret AI-assisted diagnoses, and how all of these responses recursively reorganize the diagnostic ecology. No accuracy metric captures this.
Fourth, and most fundamentally, AI diagnostic systems cannot assess mesocosmic fit. The AI reading a mammogram does not know whether the woman whose image it is examining lives in a context in which a call-back will trigger a cascade of anxiety for which her social and psychological situation provides no adequate support. It does not know whether she is the only breadwinner in a family whose stability depends on her capacity to keep working. It does not know whether she has a family history that makes any mention of cancer catastrophically recursive in her particular social world. It reads an image. The mesocosm is invisible to it.
This means that AI systems at the first threshold systematically decouple first-threshold decision from second-threshold consequence. The human radiologist, however imperfectly, could in principle integrate clinical history, patient context, and some assessment of the likely consequences of communication into the first-threshold judgment. The AI system, by design, cannot. It is optimized for accuracy with respect to a reference standard that is itself a first-threshold output -- a classification by experts of what the image shows -- rather than a second-threshold outcome: what happened to the person in whose mesocosm the classification landed.
The methodological implication is this: AI diagnostic systems need to be evaluated not only by their accuracy at the first threshold but by their recursive effects across both thresholds. What new L3 signals do they generate? Which of these are escalated into L4 categories? How do they alter radiologists' judgment and the thresholds of clinical responsibility? Do they increase the rate of second-threshold descent? What are the mesocosmic consequences of that increased descent for the populations they serve? These are not questions that can be answered by any amount of additional accuracy data. They require the conceptual vocabulary that the two-threshold model provides and that, until now, has not existed.
That the full recursive consequences of AI-assisted diagnostics cannot be predicted in advance is not a counsel of epistemic despair. It is an ontological claim about recursive systems. The consequences depend on how each agent in the system -- radiologists, clinicians, institutions, patients, families -- responds to the new situation, and those responses are not fixed in advance. What is available, and what LVT provides, is the conceptual vocabulary for tracking these consequences as they unfold: for observing what is actually happening without flattening it into a linear causal narrative, and for identifying which effects originate at the first threshold and which at the second.
VII. The Ethical Geometry of Two Thresholds
The two-threshold architecture generates an ethical geometry that is more complex than any single-threshold framework can accommodate. Two distinct sets of ethical stakes operate at two distinct sites, with different actors bearing primary responsibility at each.
At the first threshold, the primary ethical stake is epistemic responsibility: what is owed to the integrity of the biomedical stabilization process itself. The question is whether an L3 observation genuinely warrants L4 stabilization given what is known about its natural history, its rate of genuine pathology, the consequences of stabilization, and the consequences of not stabilizing. This is an ethical question with an epistemic core: it is about the accuracy and appropriateness of the classification act, considered independently of the specific mesocosm into which it may subsequently descend. The actors who bear primary responsibility here are those who make the classification: the radiologist, the pathologist, the clinician performing the screening examination, the system designers who set the detection thresholds, and the institutions that define the protocols within which these actors work.
At the second threshold, the ethical stake is different: it is mesocosmic responsibility. The question is whether a biomedical L4 category, having been produced at the first threshold, should descend into this particular mesocosm at this particular time in this particular form. This is not only, or even primarily, an epistemic question. It requires knowledge of the actual world into which the classification will land: knowledge of the patient's social circumstances, their psychological resources, their existing coordination across all five mediations, and the probable recursive consequences of descent given the specific character of that mesocosm. The actors who bear primary responsibility here are those who communicate and manage the descent: the clinician who delivers the result, frames its significance, decides what to recommend, and bears ongoing responsibility for what happens next.
The first implication of this geometry concerns the allocation of responsibility. Current medical ethics frameworks primarily address the second threshold. Informed consent, shared decision-making, the duty of candour, communication of risk -- these frameworks all concern the relationship between clinician and patient in the context of a biomedical finding that has already been made. They are second-threshold ethics. They have little to say about the ethics of the first threshold -- about the responsibility of those who decide whether an L3 observation should become an L4 stabilization in the first place. Developing an ethics of the first threshold is a task that has barely begun, and the present analysis suggests it is urgent.
The second implication concerns what might be called the trap of threshold conflation. When clinicians conflate the two thresholds -- treating the decision to stabilize and the decision to descend as a single decision, or treating the communication of a finding as automatically entailed by its detection -- they expose themselves to two distinct ethical fears simultaneously. The fear of failing to stabilize a dangerous finding drives toward maximizing detection, toward recording rather than dismissing borderline findings, toward communicating every detected abnormality. The fear of unnecessary recursive descent into patient life drives in the opposite direction. When these two fears operate without being distinguished from each other, they cannot be addressed separately, and clinical judgment becomes subject to a tension that has no principled resolution. Distinguishing the two thresholds does not dissolve this tension -- there are genuine cases in which the two ethical stakes genuinely conflict. But it clarifies what the tension is about, identifies who bears primary responsibility at each site, and makes possible a more systematic account of when the crossing of each threshold is and is not warranted.
The third implication concerns the specific situation of screening programmes. Screening programmes operate at scale: they apply first-threshold decisions to populations rather than individuals, generating biomedical L4 stabilizations for large numbers of people who have produced no L2 signal requesting clinical attention. The ethical framework that applies to individual clinical encounters -- where the encounter was initiated by a patient's felt misalignment and the clinician is responding to disruption already present -- does not straightforwardly apply. A screening programme that generates ten thousand L4 stabilizations of which eight thousand will never produce symptoms has a different ethical character from a clinical programme that generates ten thousand stabilizations of ten thousand symptomatic patients. The harm profiles are different, the obligations of disclosure are different, the requirements for informed consent are different, and the appropriate population-level weighting of individual mesocosmic consequences against population-level health benefits is a question that requires a framework capable of tracking second-threshold effects. No existing population-health ethics framework has such a capability.
A further ethical implication concerns the problem of mesocosmic invisibility. At the second threshold, the harm that descent produces is constitutively illegible within the biomedical framework that causes it. The surveillance programme that monitored Leif's prostate cancer for fifteen years records a clinical success. The fifteen years of reorganized embodiment, maintained secrecy, and anticipatory symbolic self-understanding produce no clinical finding. The system cannot subtract from what it never registered. This is not merely a measurement problem: it is an ethical problem, because harm that is constitutively invisible to an institution is harm for which that institution cannot be held accountable within its own terms. Building accountability for second-threshold harm into the institutional architecture of healthcare requires first developing the conceptual vocabulary to identify and describe that harm. The two-threshold model is a contribution toward that vocabulary.
VIII. Retrospective Intelligibility and Prospective Unpredictability
A consistent theme across the preceding sections is the claim that recursive consequences are retrospectively intelligible but prospectively unpredictable. This claim requires careful articulation, because it could be mistaken for a counsel of epistemic despair -- a recommendation that, since consequences cannot be predicted, the attempt to anticipate them should be abandoned. That is not the argument.
The claim is ontological rather than epistemic. Recursive systems are not merely difficult to predict in the way that complex physical systems are difficult to predict -- because there are too many variables, or because the data is insufficient, or because the models are underdeveloped. Recursive systems are in principle not fully predictable because their behavior is partly constituted by the states of agents who are themselves engaged in anticipatory reasoning. The person who receives a diagnosis does not simply respond to it. They interpret it, in light of what they already know and believe, in a social context that is also interpreting it, and their response is partly constituted by their anticipation of what the response will mean. The recursive loop is not merely complex. It is, in a specific sense, open.
This is why the introduction of AI diagnostic systems at the first threshold cannot be assessed through predictive modelling alone. The consequences of AI-assisted radiology for the rate of downstream diagnostic cascades, for clinical judgment, for patient anxiety, for institutional protocol, and for the eventual mesocosmic experience of the patients who receive AI-flagged results, cannot be determined in advance -- not because the models are inadequate but because the recursive consequences depend on how each agent in the system responds to the new situation, and those responses are not fixed.
Retrospective intelligibility is a different matter. Once a sequence of events has unfolded, it is possible to trace its recursive logic: to see how one threshold decision produced conditions that made another threshold decision more or less likely, how descent into a particular mesocosm reorganized coordination in ways that generated new L2 signals and new L3 articulations and new L4 stabilizations. Leif's case is retrospectively intelligible: we can trace precisely how the PSA finding reorganized each of the five mediations across fifteen years. Mrs. Saha's case is retrospectively intelligible: we can trace exactly how the ADHD classification descended into a family mesocosm and reorganized its internal dynamics in ways that produced catastrophic consequences for every member. Retrospective intelligibility is the form of understanding that is available to recursive systems, and it is the form that enables learning, accountability, and the development of better frameworks for first- and second-threshold decisions. But it is not the same as prospective prediction, and the conflation of the two is a persistent source of error in health policy.
Mesocosmic ontological literacy is the name for the capacity that retrospective intelligibility requires and that prospective analysis needs to deploy as a mode of anticipatory attention. It is not predictive capacity. It is the set of conceptual tools required to observe the unfolding of recursive consequences without flattening them into linear causal narratives, without mistaking correlation for mechanism, and without treating the absence of a measurable event as the absence of a consequence. Its practical form is the ability to ask, at each threshold, not only what the clinical evidence shows but what the recursive architecture of this condition is, which type of recursive instability it exhibits, who is likely to be affected by descent into which kinds of mesocosm, and what the probable shape of the second-threshold consequences is given the character of the condition and the character of the world into which it is descending.
This is not a capability that protocols can generate by themselves. Only the mesocosm can finally judge its own fit. But mesocosmic ontological literacy, developed as a conceptual framework and applied as a mode of clinical and research attention, can sharpen judgment at both thresholds in ways that existing frameworks cannot, because it begins from a more adequate account of what is actually happening when a clinical signal is converted into a diagnostic category and that category is allowed to descend into a living person's world.
IX. The Field in Prospect
The workshop from which this article emerged brought together medical anthropologists, clinicians, epidemiologists, and social scientists working across the Nordic welfare states. The cases presented -- AI-assisted mammography, prenatal screening, loneliness as epidemic, PMDD as diagnostic category, ADHD in context, depression across cultures, overdiagnosis in primary care , looked disparate. The two-threshold architecture reveals them as instances of the same underlying structure, located at different points in the recursive chain.
The AI mammography case concerns a transformation of the first threshold itself: what happens when a new recursive agent enters the intra-biomedical decision-making process at scale. The prenatal screening case concerns a specific form of descent into the mesocosm -- the installation of risk categories into the lives of pregnant women and their partners -- in which the second-threshold consequences are carried by a social unit rather than an individual, and in which the recursive instability is compounded by the temporal character of pregnancy as a condition of heightened anticipatory engagement. The loneliness epidemic case concerns the epidemiological categorization of a distributed social state and the consequences of that categorization's descent into daily life as a health imperative -- a case in which the first threshold (the epidemiological stabilization) and the second threshold (the descent into felt social self-assessment) are collapsed into a single media event with no intervening clinical judgment at all. The PMDD and ADHD cases concern intrinsically recursive conditions in which classification is constitutively involved in the phenomenon it names. The depression across cultures case concerns the variation in second-threshold consequences depending on the cultural and social character of the mesocosm into which biomedical L4 categories descend. The overdiagnosis case in primary care is the general case from which the two-threshold model was developed.
That all of these cases sit inside one recursive architecture is not coincidental. They are all instances of the fundamental problem this article has addressed: the failure to distinguish the decision to stabilize from the decision to descend, and the theoretical and practical consequences of that failure. The existing frameworks -- medical anthropology's focus on illness narratives and explanatory models, health economics' reliance on QALY-based outcome measurement, the medicalization critique's focus on the expansion of diagnostic authority -- have each produced genuine insights. But each begins too late, at the second threshold or at L3 and above, and each lacks the vocabulary to reach the first threshold where the recursive cascade often originates.
The overdiagnosis literature has come the closest to the empirical dimensions of the problem: it has documented, across a range of conditions and screening programmes, that rates of detection exceed rates of genuine disease, and that treatment of screen-detected conditions produces net harm in identifiable populations. What it has not yet developed is the theoretical framework for understanding why this happens at the structural level -- why the first-threshold decision is systematically biased toward stabilization, what the recursive architecture of different conditions implies for the design of screening programmes, and how the two-threshold distinction can be operationalized in research and clinical practice. The present article is a contribution toward that framework.
The most important single implication of the two-threshold model is this: overdiagnosis does not begin when a patient is told something, it begins when an intra-biomedical decision converts an observation into a category. That decision -- and the actors who make it, the institutions that structure it, the criteria that govern it, and the new AI agents that are increasingly executing it -- is the first site of accountability. No existing framework for overdiagnosis research, medical ethics, or health policy has placed that decision at the centre of its analysis. Doing so is the task that follows from what has been argued here.
Value is not what is measured, exchanged, categorized, or symbolically recognized. Value is what sustains, repairs, or improves the ongoing coordination of a living being with its world. The most valuable health states are those that never need to prove their value. The most important clinical act is sometimes the one that does not occur. And any theory of diagnostic practice that cannot account for this -- that can only see what has been stabilized, recorded, and communicated -- has already missed the most important thing. The system cannot subtract from what it never registered. The challenge of two-threshold medicine is to build frameworks capable of registering both.
References
Conrad, P. 2007. The Medicalization of Society. Baltimore: Johns Hopkins University Press.
Ecks, S. 2013. Eating Drugs: Psychopharmaceutical Pluralism in India. New York: New York University Press.
Ecks, S. 2022. Living Worth: Value and Values in Global Pharmaceutical Markets. Durham: Duke University Press.
Ecks, S. 2026. "What Medicine Cannot Register: Living Value, Verification, and the Health System's Invisible Harm." livingvaluetheory.org.
Gadamer, H.-G. 1996. The Enigma of Health. Stanford: Stanford University Press.
Heidegger, M. 1962. Being and Time. Trans. J. Macquarrie and E. Robinson. New York: Harper and Row.
Illich, I. 1976. Medical Nemesis: The Expropriation of Health. New York: Pantheon.
Jonsson, A.B.R. 2024. "Medicalization of Old Age: Experiencing Healthism and Overdiagnosis in a Nordic Welfare State." Medical Anthropology 43(4): 310-323.
Kleinman, A. 1988. The Illness Narratives. New York: Basic Books.
Welch, H.G., L. Schwartz, and S. Woloshin. 2011. Overdiagnosed: Making People Sick in the Pursuit of Health. Boston: Beacon Press.