Prescriptive Artificial Intelligence: A Formal Paradigm for Auditing Human Decisions Under Uncertainty

AI summary

Overview

Research area: Artificial intelligence, specifically decision-support systems, normative decision theory, explainable AI (XAI), imitation learning and inverse reinforcement learning, with application domains in elite sports analytics and clinical decision support.

Technical level: Advanced. The paper is built around formal axioms, separation theorems, impossibility results, and convergence statements (application of M-estimator regularity conditions, almost-sure convergence, finite-sample concentration bounds), and assumes familiarity with supervised learning, Markov decision processes, inverse reinforcement learning, and decision theory.

Scope in one sentence: The paper defines Prescriptive AI as a formally axiomatized class of systems that audit human decisions under uncertainty rather than imitate or predict them, proves two impossibility results about behavioral imitation, and demonstrates the framework's realizability in three independently engineered systems (a fuzzy soccer substitution auditor, MYCIN, and NEWS2) spanning five decades.

Note on completeness: the supplied paper text is truncated mid-sentence in Section 3.0.5. Sections describing the empirical soccer evaluation, the full GNPAF architecture (Section 4), the clinical instantiation detail (Section 5), and the discussion (Section 8.5) are referenced but their contents are not present in the provided text. Any figures from those sections are therefore reported here as not available rather than estimated.

What This Paper Is About

Most computational decision-support systems are predictive: they are trained to replicate historical human decisions and are judged by how accurately they reproduce what people did. The paper argues that this is structurally the wrong objective in high-stakes, stochastic settings, because historical decisions embed cognitive biases (status quo bias, sunk cost fallacy) and outcome bias rewards decisions that happened to succeed rather than decisions that were justified at the time. The goal is to formalize a different paradigm, Prescriptive AI, in which the system audits whether a human decision was epistemically coherent given the information available at the moment of choice, preserving human agency and accountability.

Key Contributions

1. A formal, axiomatic definition of Prescriptive AI. Four domain-independent axioms characterize prescriptive systems, and the paper gives an explicit model-theoretic semantics for the contestability axiom. Prescriptive AI is positioned as a normative subset of Prescriptive Analytics, which is itself a subset of a broader class of Action-Oriented Decision Systems. The defining criterion is the system's normative role in the decision process, not the technique it uses (symbolic, statistical, or hybrid).

2. A strengthened Imitation Incompleteness theorem. The paper proves that supervised learning from historical decisions cannot correct systematic human bias in the absence of external normative signals. The result is established under standard M-estimator regularity conditions, extended to Markovian (non-i.i.d.) decision logs, and supplemented with an explicit finite-sample concentration bound.

3. A second, independent impossibility result: Normative Non-Identifiability. The utility function underlying rational choice cannot be recovered from behavioral data under systematic bias. Unlike the classical symmetric reward ambiguity of inverse reinforcement learning (where a whole affine family of rewards is equally consistent with the data), this failure is directional: the true normative utility is provably excluded from the consistent set, and a concrete standard estimator (MaxEnt IRL) is shown to converge almost surely to a specific member that rationalizes the bias itself.

4. A role-aware cumulative mean metric and a cross-domain realizability demonstration. The metric replaces cumulative sums with a role-aware cumulative mean over fixed temporal slices, eliminating the play-time exposure bias that prevents detection of within-episode performance decay. The framework is then shown to be realized by three independently engineered systems spanning five decades: an interpretable fuzzy system for elite soccer substitution auditing, MYCIN, and NEWS2.

Main Findings

• Imitation has a structural ceiling, not a statistical one. Under standard regularity conditions, a predictor trained on historical decisions converges almost surely to the biased action rather than the normatively optimal one. Imitation performance is bounded by a structural bias term (written as ε_bias in the paper) rather than by the statistical rate O(1/√n). More data does not remove the ceiling.

• A documented predictive plateau illustrates the ceiling empirically. The paper cites prior work in which state-of-the-art substitution models plateau at approximately 70% accuracy by merely cloning human choices.

• The normative criterion cannot be learned from behavior. Normative Non-Identifiability shows the utility function behind rational choice is unrecoverable from behavioral data once the demonstrator is systematically, not merely noisily, suboptimal. The paper contrasts this directional failure with the symmetric ambiguity noted by Ng and Russell in IRL, and argues the normative criterion must be specified externally by domain authority (Axiom 2).

• Individual outcome-based labeling is ill-defined in stochastic domains. Theorem 5 formalizes an impossibility for individual-decision labeling based on binary outcomes, which limits how counterfactual and off-policy evaluation methods can be applied. The paper states this does not foreclose continuous-variable estimation or population-level calibration.

• Black-box systems cannot satisfy contestability. Theorem 7 shows that opaque systems fail the contestability axiom by construction, which is why intrinsic interpretability is treated as a structural requirement rather than a post-hoc explanation feature.

• Three structurally distinct systems satisfy the same axioms. The fuzzy soccer substitution auditor, MYCIN, and NEWS2 are presented as independently engineered instances of the six GNPAF structural components. MYCIN produced antimicrobial-therapy recommendations from an externally specified corpus of clinician-authored rules with certainty factors, was never fit to any single hospital's prescribing records, and offered WHY/HOW explanation and rule-acquisition subprograms. NEWS2 is a fixed physiological scoring protocol mandated across acute NHS trusts since 2019.

• Historical external validation of MYCIN. Yu et al. evaluated MYCIN's recommendations against nine practicing prescribers on ten real meningitis cases using a blinded panel of eight infectious-disease experts, and found MYCIN's regimens rated at least as acceptable as those of five faculty specialists.

• Human normative agreement is not recoverable by aggregation. The Moral Machine experiment, which collected millions of pairwise trolley-problem judgments across 233 countries, found systematic and culturally varying disagreement with no consensual aggregate ordering. The paper uses this as independent evidence that the normative criterion should be an externally supplied, contestable artifact rather than a statistically estimated quantity.

• Prescriptive AI is positioned as complementary to prescriptive machine learning. The paper maps Hüllermeier's five methodological challenges for learning prescriptive models onto its own axioms and theorems, arguing that the axioms characterize any system that qualifies as prescriptive, learned or not, including symbolic rule systems such as MYCIN for which a learned reward may be inapplicable.

• Decision auditing is norm-invariant backtesting. The same normative decision framework is used online (recommendation from the epistemic state at decision time) and offline (reconstruction of what would have been recommended from logged data). Future observations are used only to denoise or complete past measurements, never to introduce outcome knowledge.

Methodology in Plain English

The paper proceeds in three layers.

First, it builds a conceptual hierarchy. At the top is a broad, heterogeneous superset of "Action-Oriented Decision Systems" (optimization, simulation, rule-based expert systems, reinforcement learning agents, recommenders, and predictive decision support). Inside that sits Prescriptive Analytics, which is decision-oriented but normatively under-specified. Inside that sits Prescriptive AI, defined by four axioms: it must handle uncertainty and asymmetric risk explicitly, its normative criterion must be externally specified rather than fit to biased behavioral data, it must be intrinsically interpretable, and its recommendations must be contestable so that human agency is preserved. A system that cannot be inspected, audited, or contested can only suggest or automate; it cannot prescribe.

Second, it proves separation and impossibility results. The argument is that a predictor trained by supervised learning to map contexts to historical decisions converges to the historical policy by construction, so any systematic bias in the historical data survives asymptotically. This is contrasted with offline reinforcement learning, which reintroduces outcome bias because the reward signal is derived from realized outcomes, and with inverse reinforcement learning, which assumes an approximately optimal demonstrator, an assumption that fails exactly when bias is systematic. The paper leans on van Benthem's account of decision-making as an epistemic state transition and on Savage's axiomatization of rational choice to argue that decision quality is defined at time t, before the outcome at time t+n is realized, and it gives a formal correspondence (Proposition 4) showing a Prescriptive AI recommendation of the form argmax over actions of expected utility realizes the ex-ante decision rule that Savage's representation theorem attributes to a rational agent given (U, P).

Third, it demonstrates realizability rather than proposing a new algorithm. The soccer substitution auditor is a hybrid statistical-symbolic system that overlays intrinsically interpretable fuzzy reasoning onto statistical signals, treating the output as an auditing layer that surfaces latent risk such as performance decay or defensive liability before it shows up as observable failure. Because a single domain could produce a domain-specific artifact, the paper checks whether the same axiomatic structure is satisfied by MYCIN and NEWS2, neither of which was engineered with this framework in mind.

Why This Matters

Impact on research. The paper reframes a practical engineering question ("how do we build better decision-support models?") as a normative one ("what must any system satisfy to count as prescriptive?"). If the Imitation Incompleteness and Normative Non-Identifiability results hold, they set a limit on a large family of methods, including behavioral cloning, offline RL, and inverse reinforcement learning, when the demonstrator is systematically biased. That shifts the research target from learning algorithms to the external specification, inspection, and contestation of normative criteria, and gives a common vocabulary for comparing symbolic clinical rule systems with modern statistical decision support.

Real-world applications.

• Elite sports (soccer substitutions). Time-sensitive, irreversible tactical decisions made under pervasive uncertainty, where auditing can surface performance decay and defensive risk that cumulative playing-time metrics and outcome-based judgment obscure.
• Clinical triage and early warning. The paper's own clinical examples, MYCIN for antimicrobial therapy recommendation and NEWS2 for early in-hospital mortality risk, illustrate rule bases that are externally specified and inspectable by the treating physician.
• Operational risk management. Settings with irreversible actions and asymmetric consequences, where the question is whether a decision was justified by the evidence available at the time rather than whether it happened to work out.
• Strategic resource allocation. Allocation decisions where institutional inertia and status quo bias tend to make predictive systems validate conservative behavior rather than reveal when deviation is warranted.

Industry relevance. Organizations deploying AI decision support in regulated or safety-critical contexts face two recurring problems: black-box recommendations invite automation bias, and outcome-based performance reporting makes it impossible to tell a good decision from a lucky one. The paper's framework gives a structure for building systems whose reasoning is transparent by design and whose recommendations can be challenged, and it argues that the normative criterion should be an explicit, inspectable, contestable artifact supplied by domain authority rather than inferred from the organization's own historical decisions.

Future Directions

1. Quantify the bias term in deployed settings. The Imitation Incompleteness theorem bounds imitation performance by ε_bias. The paper does not report empirical estimates of that term for the soccer or clinical instantiations in the provided text, leaving open how large it is in practice and how to measure it.

2. Study richer feedback and partially learnable components. The paper positions counterfactual risk minimization and doubly robust estimators as legitimate ways to implement specific GNPAF components when richer logged (s, a, r) feedback is available, rather than as substitutes for an externally specified normative criterion. How much of a prescriptive system can safely be learned, and which components cannot, is left as an open design question.

3. Extend beyond single-agent individual decision-making. The formalization is explicitly scoped to high-stakes, stochastic settings involving single-agent individual decision-making. Multi-agent and team decision settings, where the soccer domain already hints at role-aware evaluation across interacting players, are not covered.

4. Test external specification as a practical process. The paper argues the normative criterion should be supplied externally and made explicit and contestable, citing the failure mode of Asimov's Three Laws (natural-language rules admitting conflicting or under-determined interpretations at the point of application) as an instance of the same specification problem. How domain authorities should author, version, and adjudicate such criteria remains unresolved.

Target Audience

This paper is most useful to AI researchers working on decision support, imitation learning, offline reinforcement learning, inverse reinforcement learning, and explainable AI who want a formal account of why imitation-based systems cannot correct systematic bias. It also speaks to clinical informatics and sports analytics practitioners who need interpretable, contestable decision tools, and to policy and governance audiences concerned with how normative criteria for automated decision support should be specified, audited, and held accountable. Readers without a background in decision theory, measure-theoretic convergence results, or reinforcement learning will find the theoretical sections demanding, and may prefer the taxonomy (Section 3) and the realizability discussion of MYCIN and NEWS2 (Section 5).

Authors’ abstract

We formalize Prescriptive Artificial Intelligence as a distinct paradigm for human-AI decision collaboration in high-stakes, stochastic environments involving single-agent individual decision-making. Unlike predictive systems optimized for outcome accuracy, prescriptive systems audit human decisions under uncertainty, providing normative guidance while preserving human agency and accountability. We introduce four domain-independent axioms characterizing prescriptive systems and prove fundamental separation results. Central is the Imitation Incompleteness theorem: supervised learning from historical decisions cannot correct systematic biases in the absence of external normative signals. Under standard regularity conditions, the induced predictor converges almost surely to the biased action rather than the normatively optimal one. Performance in decision imitation is therefore bounded by a structural bias term (epsilon_bias) rather than the statistical rate O(1/sqrt(n)), a result extended to Markovian logs and finite-sample concentration bounds. We demonstrate realizability through three independent instantiations spanning five decades: an interpretable fuzzy system for elite soccer auditing, revealing decision latency and risk states obscured by outcome and status quo biases; MYCIN, the historically validated rule-based clinical consultation system; and NEWS2, a nationally mandated clinical protocol validated on a prospective multi-center cohort. The framework establishes Prescriptive AI as a general, realizable class of decision-support systems for safety-critical domains where interpretability, contestability, and normative alignment are essential.