From the contestable decision to the default trajectory: directional neutrality and terminal legitimation of algorithmic care
Three independent events converged in May 2026. In the Estate of Lokken v. UnitedHealth Group litigation, a Minnesota court ordered in March the production of documents: the complaint targets the nH Predict model, developed by naviHealth, to which is attributed an error rate of approximately 90 percent in post-acute care denials for Medicare Advantage beneficiaries. On 19 May, health reporting established that human review of clinical context had been almost entirely delegated to automated systems. On 21 May, HHS announced an intensification of AI use to track fraud in federal health expenditure.
The convergence is not fortuitous. The same class of predictive tool, applied to a healthcare expenditure, serves at the insurer to compress care access and at the public regulator to detect fraud. The debate that followed distributed moral judgments: one use abusive, the other virtuous. That distribution misses the structural property that explains both: the calculation does not know which end it serves. It optimizes a perimeter.
A coverage management model does not predict the patient’s condition. It predicts a cost trajectory and calibrates care against that trajectory. The distinction is categorical and has a name: the optimization perimeter.
Three perimeters compete over the same decision. The clinical perimeter maximizes patient benefit and minimizes loss of chance. The budgetary perimeter minimizes avoidable expenditure. The capacity perimeter absorbs flow and prevents saturation. None is inherently illegitimate. The wrong arises from one thing: a system that claims to operate within the clinical perimeter while actually optimizing a budgetary or capacity one. This is not a performance failure, it is a categorical error on the objective, a false ontology of purpose. The system is not mistaken within its perimeter; it operates within a different perimeter than the one attributed to it.
The morality of the use case resides in the institution that assigns the perimeter, never in the model. A governance device that monitors the model without monitoring the assignment of the perimeter is watching the wrong variable. This holds equally for the insurer that denies post-acute care and for the regulator that hunts fraud: what differs is the assigned perimeter, not the architecture.
UnitedHealth’s defense, to the effect that coverage decisions are made by medical directors and not by the AI, is accurate in its letter and misleading in its scope. This formulation is the most important to examine, because it is the one most likely to satisfy a regulatory inquiry without resolving the structural problem.
The relevant question is not who signs. It is who composes the perimeter on which the signature bears. If a medical reviewer validates in bulk a recommendation produced on the cost perimeter, the signature does not reintroduce the clinical perimeter: it authenticates its absence. The human is then in the loop without power over the loop. Present, but without perimeter.
This distinction, human-in-the-loop versus human-as-alibi, conditions the entire regulatory discussion that follows. A requirement of human oversight that does not specify which perimeter the human is overseeing does not add governance to the system. It adds signature to a system that already produces it.
The three-generation ordering clarifies what is actually at stake. The first generation is the explicit human refusal: a dated, signed, situated decision, contestable on its own terms. The second is the algorithmically recommended decision validated by a human: the model proposes, the human disposes in principle, and a specific case file remains identifiable and attackable. Both generations share the same object: a decision on a dossier.
The third generation changes the object entirely. It no longer optimizes a decision; it optimizes a decision space. A generation 3 system acts by distributed modification of the conditions of access, priority, friction, or capacity that structure the terrain upstream of the local clinical decision. It does not reject the case file: it deforms the terrain on which the case file will be processed. When a denial occurs, it is no longer a cause but the side effect of an already inclined trajectory.
This is the structural consequence that formulations centered on automated decisions miss. Triage tools, scoring systems, and capacity instruments do not always decide on a patient; they transform the context in which that patient’s deprivation becomes probable. The wrong does not reside in an act but in a default trajectory. The progression from decision to trajectory is the progression from the contestable to the near-invisible: we know how to challenge a denial; we do not know how to challenge a probability.
Generation 3 remains a structural hypothesis. Its falsifier is known and measurable: the human override rate, by class of case. That data is publicly available nowhere. Its absence of publication is itself a signal of poor governability.
The same predictive model applied to healthcare expenditure serves, at the Assurance Maladie, to detect and halt fraud (723 million euros stopped in 2025, up 15 percent, through a decade of datamining); at the CPAM de Paris, to assign an alert level to each case file since August 2025; and, if the Lokken allegations prevail, to compress post-acute care access at the patient’s expense.
Directional neutrality names this architectural property: the same artifact can be reoriented toward opposite ends without changing its technical grammar. It does not designate an axiological neutrality of the model, the false idea that AI is without values. It designates instrumental invariance: what changes is the perimeter assigned by the institution, not the structure of the calculation. The architecture is commutable; the purpose is not.
The danger is not that AI is neutral. The danger is that it is reusable, and that this reusability passes unnoticed behind the apparent morality of the use case. An insurer that assigns a budgetary perimeter and a regulator that assigns a fraud-detection perimeter are making opposite institutional choices with the same technical instrument. Both choices are principally governed by the perimeter decision, not by model quality. Directing regulatory attention toward model audits without directing it toward perimeter declarations leaves the decisive variable unobserved.
A European executive might close the file at this point: American matter, private insurers, solidary systems are different. This framing error must be closed immediately.
Europe does not need prior authorization to encounter the problem. It needs only to modulate access, delay, intensity, priority, and the administrative burden of the care pathway. The second distinction that cuts: denial is one modality of deprivation; delay is another; friction is a third. Modern deprivation rarely passes through an explicit no. It passes through the document request that extends the timeline, the low-priority routing that delays the appointment, the enhanced fraud-scoring threshold that triggers additional review, the SLA weighting that guarantees shorter timelines for some clinical profiles than others.
The French terrain is already there without anything spectacular. Fraud scoring sorts case files by alert level. Capacity management optimizes flows: the Calyps algorithm has predicted activity at the Valenciennes hospital center since 2021 with approximately 95 percent reliability at 48 hours. Queue prioritization, BPM orchestration of case file steps, admission control on queues, dynamic documentary control thresholds, claim workflow scoring: each of these calibrates a condition of access without any clinical decision being formally made, and therefore without any formally being contestable. France is not yet industrializing explicit algorithmic refusal; it is already industrializing algorithmic modulation of the care pathway.
The central dissociation: a system can remain locally compliant, with each routing rule, threshold, and SLA meeting its specification, while globally producing an attrition of access to care that no decision has ordered. Compliance is verified at the component level; attrition occurs at the system level. Neither contradicts the other, and that is precisely what makes them formidable in combination.
The objection from advocates of quality-based regulation runs as follows: audit the model, impose effective human review, and the device becomes governable again. This objection contains a share of truth: if review genuinely reassigned the decision to its clinical perimeter, the thesis would fall. But it presupposes that convergence between human and model would be a failure, a laziness, an alibi correctable through discipline. That convergence is an expected property of the work system, not a fault of the reviewers.
The human converges with the model by design. The structural factors are cumulative: limited time per case file, informational asymmetry facing a system the reviewer did not build, productivity pressure, reluctance to deviate from a traced recommendation, absence of access to the model’s causal reasoning, responsibility diffused across the chain. Most determinatively: override cost exceeds validation cost. Overriding requires time, justification, personal exposure; validating requires none of these. A system that makes overriding more costly than validating mechanically produces validation. The human signature ceases to be a control and becomes a terminal legitimation, not through vice but through structure.
“More review” does not suffice: it adds signature to a system that already produces it. What is missing is the measurement of effectivity and the reattribution of the perimeter. Five requirements would make the device actionable: declare the perimeter actually optimized (clinical, economic, capacity, or fraud-detection); measure the actual human override rate by class of case; trace causal transitions from score to alert to review to decision to pathway effect; identify nominative responsibility for each critical transition; publish or audit false positive and false negative classes with an effect on care access.
The regulatory framework confirms the stakes through its own prudence. Article 14 of the AI Act imposes human oversight for high-risk systems but its effectivity remains to be operationalized. The Digital Omnibus compromise of 7 May 2026 is set to defer high-risk obligations to December 2027 for autonomous systems and August 2028 for embedded systems. This interregnum is precisely the space where facade review flourishes and dissociation becomes a viable strategy.
The operational requirement is a perimeter requirement before it is a performance requirement. The instruction is not “audit your models.” It is: declare the perimeter on which the care trajectory is calculated, and measure how frequently a human departs from it.
A contrast illustrates the point. PREDICARE, within the territorial predictive medicine program, is a predictive device oriented toward patient decompensation. Its optimization perimeter is physiological: anticipation window, alert threshold, moment of intervention. Its governance difficulty is intact, but it is the right perimeter: the error one fears is an error on the patient’s condition, contestable by the clinician on their own terrain. A coverage or regulatory system with comparable predictive structure bears on the cost or capacity trajectory. Same form, inverse perimeter. The prediction inherits from its perimeter. Governing a predictive system begins with governing what it predicts the trajectory of.
Care no longer needs to be explicitly refused; it can be rendered statistically improbable by a decision environment optimized for a perimeter other than clinical benefit. When the wrong has dissolved into a trajectory, the last point of leverage is neither the decision, which did not occur, nor the signature, which was only a legitimation: it is the initial condition. The assigned perimeter, the calibrated threshold, the routing rule, the encoded priority, everything that inclines the terrain before the first case file enters it.
Governing algorithmic care is no longer a matter of arbitrating denials. It is a matter of reclaiming the initial conditions before they become a trajectory that no one will know how to contest: not the patient, who has nothing to challenge; not the clinician, who signed nothing; not the institution, which only assigned a perimeter.
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.