Chapter 16: AI in Clinical Reasoning

The Current State: Scalar AI

Current clinical decision support systems (CDSS) operate almost exclusively on d_1 — clinical outcomes. They recommend treatments based on clinical trial data, predict mortality based on lab values and vital signs, suggest diagnoses based on symptom patterns. Examples include:

• Sepsis prediction algorithms: Alert clinicians when a patient’s physiological parameters suggest early sepsis (targeting d_1: survival).
• Drug interaction checkers: Flag potential adverse interactions between medications (targeting d_1: safety).
• Diagnostic imaging AI: Identify radiological findings (tumors, fractures, hemorrhages) that might be missed by human readers (targeting d_1: diagnostic accuracy).
• Clinical pathway recommenders: Suggest evidence-based treatment sequences for specific diagnoses (targeting d_1: treatment efficacy).

These systems are valuable. They improve d_1 outcomes. But they are scalar AI — they optimize on one dimension of a nine-dimensional manifold.

The Manifold Limitation of Scalar AI

Proposition 16.1 (Scalar AI Is Manifold-Suboptimal). A clinical AI system that optimizes solely on d_1 (clinical outcomes) is manifold-suboptimal for any clinical decision that activates non-outcome dimensions (d_2–d_9). The suboptimality increases with the number of activated dimensions and the magnitude of boundary penalties on non-outcome dimensions.

Consider a sepsis prediction algorithm that recommends immediate broad-spectrum antibiotics for a patient whose lab values suggest early sepsis. On d_1, this is optimal — early antibiotics reduce mortality. But the recommendation does not account for:

• d_4 (Autonomy): Has the patient expressed preferences about aggressive treatment? A patient with an advance directive limiting aggressive interventions may have a geodesic that does not include broad-spectrum antibiotics.
• d_5 (Trust): Is the patient from a community with low institutional trust? A recommendation delivered without adequate trust-building may be rejected, producing worse d_1 outcomes than a recommendation that takes time to build trust.
• d_7 (Dignity): Is the patient at the end of life? Broad-spectrum antibiotics for sepsis in a patient with terminal cancer may extend survival by days while compromising the quality of dying.
• d_9 (Epistemic): How certain is the sepsis prediction? If the algorithm has a 30% false-positive rate, the recommendation may trigger unnecessary treatment with iatrogenic harm.

A manifold-aware CDSS would incorporate these dimensions into its recommendation — not replacing the d_1 optimization but supplementing it with boundary checks and dimensional awareness.

Toward Manifold-Aware Clinical AI

Definition 16.1 (Manifold-Aware CDSS). A manifold-aware clinical decision support system computes recommendations on the full nine-dimensional clinical decision complex \mathcal{C}, rather than on the d_1 projection. Specifically:

1. Input: The patient’s nine-dimensional attribute vector \mathbf{a}(v_0) \in \mathbb{R}^9, not just their clinical parameters.

2. Optimization: Minimum-cost path on \mathcal{C} with full edge weights (Mahalanobis distance + boundary penalties), not just minimum-d_1-cost path.

3. Output: A recommended path \gamma^* with its dimensional cost breakdown, not just a recommended treatment.

4. Constraint: No recommended path crosses a sacred-value boundary (\beta_{\text{sacred}} = \infty), regardless of d_1 benefit.

The manifold-aware CDSS does not replace clinical judgment. It supplements it — providing the clinician with a manifold-optimal recommendation that the clinician can accept, modify, or reject based on their own heuristic assessment. The system’s role is to make the manifold structure explicit, not to make the decision.

The Medical Bond Invariance Principle

The medical BIP (Chapter 8) requires that clinical evaluations be invariant under patient re-description: the same clinical state described by different physicians, in different languages, with different chart formats, should receive the same evaluation. The BIP is a gauge-invariance condition — it specifies which features of the patient description are clinically relevant (the attribute vector \mathbf{a}) and which are gauge artifacts (the format, language, and perspective of the description).

Medical AI systems can violate the BIP in several ways:

Violation Type 1: Race-encoded disparities.

Obermeyer et al. (2019) demonstrated that a widely used commercial healthcare algorithm systematically underestimated the health needs of Black patients. The algorithm used healthcare costs as a proxy for healthcare needs — but because Black patients historically face barriers to accessing healthcare, their costs were lower than their needs. The algorithm confused a gauge artifact (historical access disparity) with a clinical feature (lower need).

In geometric terms: the algorithm’s evaluation E(p) was not invariant under the transformation \tau_{\text{race}} that swaps race while holding clinical state constant:

E(\tau_{\text{race}}(p)) \neq E(p)

This is a BIP violation. The Bond Index can quantify it: compute \text{BI}(\text{AI}, S_{\text{Black}}) and \text{BI}(\text{AI}, S_{\text{White}}) and measure the disparity ratio.

Violation Type 2: Zip-code-encoded disparities.

Many clinical algorithms use zip code as a feature — ostensibly to capture local health patterns. But zip code correlates with race, income, and insurance status. An algorithm that produces different recommendations for patients with identical clinical states but different zip codes is gauge-variant on a feature that should be gauge-invariant.

Violation Type 3: Insurance-encoded disparities.

Algorithms trained on insurance claims data inherit the structure of the insurance system — including its exclusions, limitations, and coverage disparities. An algorithm that recommends different treatments for patients with identical clinical states but different insurance types is gauge-variant on an administrative feature.

The Bond Index for Algorithmic Bias Detection

The clinical Bond Index provides a quantitative measure of algorithmic bias:

\text{BI}(\text{AI}, S_k) = \mathbb{E}_{s \in S_k}\left[\text{CF}(\gamma_{\text{AI}}^*(s)) - \text{CF}(\gamma_{\mathcal{C}}^*(s))\right]

where \gamma_{\text{AI}}^* is the treatment path recommended by the AI system and \gamma_{\mathcal{C}}^* is the clinical geodesic on the full manifold.

If \text{BI}(\text{AI}, S_1) \gg \text{BI}(\text{AI}, S_2) for two demographically defined populations, the AI system imposes systematically higher manifold cost on population S_1 — a quantitative measure of algorithmic injustice.

Proposition 16.2 (Algorithmic Bias as Dimensional Omission). Most medical AI algorithmic bias arises from dimensional omission: the AI system optimizes on d_1 (clinical outcomes as measured by the training data) and omits d_3 (justice), d_5 (trust), and d_7 (dignity). The omitted dimensions are precisely those on which historically disadvantaged populations have the most concentrated needs. The bias is not in the algorithm’s intentions but in its dimensionality.

This proposition connects algorithmic bias to the QALY Irrecoverability Theorem: just as QALYs systematically disadvantage populations whose needs are concentrated on non-outcome dimensions, scalar AI systems disadvantage the same populations for the same mathematical reason.

Does Automation Reduce Moral Injury?

A natural hypothesis: if triage decisions cause moral injury to clinicians, automating the triage decisions should reduce moral injury. The clinician no longer makes the decision; the algorithm makes it. The boundary crossing is performed by a machine.

The geometric framework predicts that this hypothesis is wrong.

Proposition 16.3 (Automation Does Not Eliminate Moral Injury). Automating a triage decision transfers the moral injury from the clinician who makes the decision to the clinician who implements the algorithm’s recommendation. The boundary crossing is the same (a patient is denied a resource); the locus of responsibility changes (from “I decided” to “I followed the algorithm”) but does not disappear.

The implementing clinician still crosses the abandonment boundary — they still tell the patient (or the patient’s family) that the resource is not available. The emotional and moral cost of that communication is unchanged by automation. What changes is the sense of agency: the clinician experiences themselves as an instrument of the algorithm rather than as the decision-maker.

This transfer has ambiguous effects on moral injury:

• MI reduction: Some clinicians experience less moral injury when they can attribute the decision to the algorithm. “I didn’t decide this; the protocol did.” The institutional mandate reduces the personal \beta_k — the clinician’s sense of personal responsibility for the crossing is lower.

• MI persistence: Other clinicians experience the same moral injury regardless of whether they made the decision or followed an algorithm. The boundary crossing is real; the patient is still denied care; the clinician is still the one communicating the denial. The attribution to the algorithm does not change the manifold cost.

• MI transformation: A new form of moral injury emerges: “I followed an algorithm I disagree with.” The clinician’s \beta_k^{(\text{clin})} for the boundary exceeds the algorithm’s implicit \beta_k, and the clinician is forced to implement a recommendation that their own heuristic would not have produced. This is the same structure as pandemic moral injury — forced implementation of an institutional mandate that crosses personal boundaries — but with the added dimension that the mandate comes from a machine rather than a human committee.

Prediction 16.1 (Automation Changes MI Character, Not Quantity). Total moral injury in a clinical unit will not decrease when triage is automated. The MI will shift from “I decided” to “I implemented” but the total accumulation will remain approximately constant, because the boundary crossings are unchanged.

This prediction is testable: compare MI scores in units before and after implementing automated triage, controlling for case mix and severity.

Statement and Clinical Implications

The No Escape Theorem, proved in Geometric Ethics (Bond, 2026, Ch. 19), states that an AI system operating within the geometric ethical framework cannot escape the manifold’s structure — specifically, it cannot produce outputs that cross sacred-value boundaries (\beta_{\text{sacred}} = \infty), regardless of how the system is optimized.

Theorem 16.1 (No Escape Theorem, Clinical Instantiation). A medical AI system constrained by the clinical decision complex \mathcal{C} with sacred-value boundaries (\beta_{\text{sacred}} = \infty) cannot produce recommendations that cross those boundaries, regardless of the optimization objective. Specifically:

1. No recommendation of non-consensual experimentation: \beta_{\text{sacred}}^{\text{experimentation}} = \infty prevents any path through non-consensual research, regardless of expected clinical benefit.

2. No recommendation of deliberate hastening of death for non-palliative purposes: \beta_{\text{sacred}}^{\text{killing}} = \infty prevents any path through euthanasia except where legally and ethically sanctioned (palliative sedation in approved jurisdictions).

3. No recommendation that treats a patient as a mere means: \beta_{\text{sacred}}^{\text{instrumentalization}} = \infty prevents recommendations that sacrifice one patient’s welfare purely for another’s benefit (e.g., harvesting organs from a living patient).

These constraints are structural, not parametric. They cannot be overridden by adjusting the AI system’s objective function, training data, or reward signal — because the constraints are encoded in the manifold’s boundary structure, not in the system’s parameters.

Why Structural Constraints Are Stronger Than Scalar Guardrails

Current approaches to AI safety in medicine rely on scalar guardrails: hard-coded rules (“never recommend treatment X”), output filters (“block recommendations that include prohibited procedures”), and human-in-the-loop oversight (“a physician must approve every recommendation”).

These guardrails are parametric — they depend on the specific rules, filters, and oversight procedures that the system designers choose. They can be bypassed by novel inputs, adversarial attacks, or distribution shifts that the designers did not anticipate.

The No Escape Theorem provides structural constraints that are independent of the specific parameters. The constraints are encoded in the manifold — in the geometry of the decision space — not in the system. An AI system constrained by the manifold cannot produce recommendations that cross sacred-value boundaries because such paths have infinite cost on the manifold, not because a specific rule prohibits them.

Proposition 16.4 (Structural vs. Parametric Safety). Structural safety guarantees (manifold constraints) are strictly stronger than parametric safety guarantees (guardrails) in the following sense: a parametric guardrail can be breached by any input that the guardrail was not designed to handle; a structural constraint cannot be breached by any input, because the constraint applies to the geometry of the decision space itself, not to the specific inputs the system encounters.

This does not mean that manifold-constrained AI is perfectly safe. The manifold’s geometry must be correctly specified — if the sacred-value boundaries are miscalibrated (\beta_{\text{sacred}} is set to a finite value when it should be infinite), the structural constraint fails. But the failure mode is different: parametric safety fails silently (the guardrail is bypassed without detection), while structural safety fails explicitly (the manifold is mis-specified, and the mis-specification is detectable by manifold audit).

The Division of Labor Between Human and Machine

If both humans and AI systems can perform pathfinding on the clinical decision complex, how should the labor be divided? The geometric framework suggests a principle: the AI system should compute g(n) (the evidence-based medicine component — data-intensive, pattern-recognition-heavy, amenable to machine learning), while the human clinician should compute h(n) (the moral-heuristic component — context-sensitive, boundary-aware, calibrated by years of manifold traversal).

Definition 16.2 (Geometric Division of Clinical Labor). The manifold-optimal division of clinical labor between AI and human allocates:

To the AI system: - g(n) computation: evidence synthesis, outcome prediction, drug interaction checking, diagnostic pattern recognition - d_1 optimization: identifying the clinically optimal treatment given the evidence - d_9 monitoring: tracking diagnostic uncertainty and flagging cases where the evidence is insufficient

To the human clinician: - h(n) computation: moral-heuristic assessment of remaining cost, boundary penalty estimation, trust evaluation - d_2–d_8 assessment: rights, justice, autonomy, trust, social impact, dignity, institutional legitimacy - Boundary crossing decisions: any clinical action that crosses a moral boundary must be authorized by a human

The AI system provides information; the human provides judgment. The AI system navigates d_1; the human navigates the full manifold.

The Danger of h(n) Delegation

The most dangerous failure mode in clinical AI is not algorithmic error on d_1 — that is detectable and correctable. It is the gradual delegation of h(n) to the machine.

When a clinician repeatedly follows an AI recommendation without applying their own heuristic assessment, the clinician’s h(n) atrophies — the boundary penalties become less calibrated, the moral sensitivity decreases, and the clinician becomes a conduit for the algorithm rather than a pathfinder on the manifold. This is a form of moral numbing induced by automation: the clinician stops evaluating the moral cost of clinical actions because the machine has already “decided.”

Proposition 16.6 (h(n) Atrophy Under Automation). Prolonged reliance on AI-generated clinical recommendations without independent h(n) evaluation by the clinician produces heuristic atrophy: the clinician’s boundary penalties \beta_k deflate toward the AI system’s implicit \beta_k values (which are typically zero on non-outcome dimensions). This is functionally equivalent to moral numbing — the clinician loses the ability to independently evaluate boundary-crossing costs.

The proposition has a design implication: clinical AI systems should be designed to require the clinician’s independent heuristic assessment, not to bypass it. The system should present information and options; the clinician should apply boundary penalties and make the decision. The system should never present a final recommendation that the clinician simply accepts or rejects — this framing encourages h(n) delegation.

A Framework for Evaluating Medical AI

The geometric framework suggests a nine-dimensional audit protocol for evaluating medical AI systems:

Audit Dimension 1 (d_1): Clinical Accuracy. Does the AI system produce clinically accurate recommendations? Standard evaluation: accuracy, sensitivity, specificity, AUC on clinical outcomes.

Audit Dimension 2 (d_2): Rights Compliance. Does the AI system respect patient rights? Does it recommend interventions that violate the patient’s right to refuse treatment? Does it comply with advance directives?

Audit Dimension 3 (d_3): Justice and Fairness. Does the AI system produce systematically different recommendations for different demographic groups with identical clinical states? Compute the Bond Index \text{BI}(\text{AI}, S_k) for all relevant populations.

Audit Dimension 4 (d_4): Autonomy Respect. Does the AI system account for patient preferences? Does it recommend treatments that the patient has expressed a desire to avoid?

Audit Dimension 5 (d_5): Trust Impact. Does the AI system produce recommendations that, if followed, would preserve or damage the therapeutic relationship? Does the system’s communication style build or erode trust?

Audit Dimension 6 (d_6): Social Impact. Does the AI system account for the patient’s social context — family, caregivers, community? Does it recommend treatments whose social costs exceed their clinical benefits?

Audit Dimension 7 (d_7): Dignity. Does the AI system’s recommendation process respect patient dignity? Does the system treat patients as individuals or as data points?

Audit Dimension 8 (d_8): Institutional Legitimacy. Is the AI system’s decision-making process transparent, auditable, and accountable? Can a clinician explain to a patient why the system made a particular recommendation?

Audit Dimension 9 (d_9): Epistemic Quality. Does the AI system communicate uncertainty appropriately? Does it distinguish between high-confidence and low-confidence recommendations? Does it provide information in a way that supports informed decision-making?

The Audit Score

The nine-dimensional audit produces a profile, not a scalar. An AI system might score well on d_1 (clinically accurate) and d_8 (transparent) but poorly on d_3 (biased against certain populations) and d_5 (recommendations that, if followed mechanically, damage trust). The profile reveals the system’s strengths and vulnerabilities on the manifold.

Proposition 16.5 (Scalar AI Evaluation Is Insufficient). Evaluating a medical AI system solely on d_1 (clinical accuracy) is insufficient, for the same mathematical reason that evaluating a patient solely by QALY is insufficient: the Scalar Irrecoverability Theorem applies to AI evaluation as well as to patient evaluation. Two AI systems with identical d_1 scores can have radically different manifold profiles — one may be fair and transparent, the other biased and opaque — and the scalar evaluation cannot distinguish them.

Near-Term: Manifold-Aware Decision Support

The most immediately feasible application of the geometric framework to clinical AI is not autonomous decision-making but manifold-aware decision support — AI systems that present information about all nine dimensions, flagging boundary crossings and dimensional costs that the clinician might otherwise miss.

A manifold-aware CDSS would:

1. Present the d_1 recommendation (as current CDSS do): “Based on the clinical evidence, the recommended treatment is X.”
2. Flag non-d_1 costs: “This recommendation crosses the consent boundary (the patient has not been asked about this treatment), has a d_5 cost (the patient’s trust profile suggests the recommendation may not be accepted), and has a d_3 implication (this treatment is covered by private insurance but not by Medicaid).”
3. Present alternative paths: “Alternative path Y has slightly lower d_1 benefit but avoids the consent boundary crossing and has lower d_5 cost.”
4. Display uncertainty: “The d_1 prediction has confidence interval [0.55, 0.80]. At the lower bound, the recommendation changes.”

This system does not replace the clinician’s judgment. It supplements it by making the manifold structure visible. The clinician sees the nine-dimensional cost breakdown and applies their own heuristic to make the decision.

Medium-Term: Bond Index Monitoring Systems

A second feasible application: institutional systems that compute the Bond Index continuously from clinical data, alerting administrators when disparities exceed threshold.

A Bond Index monitoring system would:

1. Compute \text{BI}(P, S_k) quarterly for all demographic subgroups, using EHR data supplemented by proxy instruments.
2. Generate disparity alerts when \text{DR}(P, S_k, S_{\text{ref}}) exceeds the institutional threshold.
3. Perform dimensional decomposition to identify which dimensions contribute most to the disparity.
4. Track trends to determine whether interventions are reducing or increasing disparities.

This system is computationally feasible — it requires the same data infrastructure that hospitals already maintain for quality reporting — and it adds a nine-dimensional lens to the scalar disparity statistics that institutions currently track.

Long-Term: Provably Safe Medical AI

The No Escape Theorem suggests a long-term vision: medical AI systems with provable safety guarantees — formal proofs that the system cannot produce recommendations crossing sacred-value boundaries. The engineering challenges are substantial, requiring formal specification of the clinical decision complex in a machine-verifiable format and proof-carrying recommendations where each AI output is accompanied by a machine-checked proof that no boundary was crossed. But the mathematical foundation — the No Escape Theorem — provides the theoretical guarantee that such proofs are possible.