Chapter 10: Decision Theory, Analytics and Artificial Intelligence

1. Introduction

Information technology increasingly functions as a decision-making infrastructure. Digital systems do not merely store transactions or automate routines; they classify risk, allocate resources, recommend actions, forecast events, generate content, and influence human judgment. Consequently, advanced information technology requires more than technical proficiency. It requires an integrated understanding of statistical inference, Bayesian reasoning, optimization, utility, causal analysis, machine learning, and the institutional conditions under which evidence becomes a decision.

Decision theory provides a formal language for selecting actions under certainty, risk, and uncertainty. Analytics transforms data into descriptive, predictive, and prescriptive insight. Artificial intelligence (AI) extends this capability through computational models that learn patterns, reason over representations, and increasingly generate text, images, code, and decisions. These fields are complementary but not interchangeable. A highly accurate predictive model may be unsuitable for a policy decision if it relies on unacceptable data, creates inequitable outcomes, or does not identify causal effects. Similarly, a generative AI system may produce fluent content without providing reliable evidence.

This chapter develops a critical framework for integrating these disciplines. It emphasizes that good decision making is not simply a matter of choosing the most sophisticated algorithm. It involves defining objectives, representing uncertainty, identifying consequences, distinguishing correlation from causation, validating models, governing their use, and maintaining meaningful human responsibility.

2. Learning Objectives

After studying this chapter, the reader should be able to:

1. Explain the relationship among decision theory, analytics, causal inference, optimization, and AI.

2. Distinguish decisions under certainty, risk, and uncertainty.

3. Apply expected utility, Bayesian updating, and value-of-information reasoning.

4. Explain descriptive, diagnostic, predictive, and prescriptive analytics.

5. Critically compare statistical and machine-learning models.

6. Distinguish predictive association from causal effect.

7. Evaluate explainability, fairness, robustness, and calibration in AI systems.

8. Design an evidence-based human–AI decision process.

9. Assess generative AI systems in terms of utility, reliability, governance, and risk.

10. Formulate managerial and research implications for responsible AI adoption.

3. Meaning and Concept

Decision theory studies how an agent should choose among available actions when outcomes depend partly on uncertain states of the world. A decision problem can be represented by actions, states, consequences, probabilities, and preferences. For example, an organization deciding whether to expand a data centre may compare actions such as expand, delay, or outsource; states such as high or low demand; and consequences such as profit, cost, resilience, and strategic flexibility.

Analytics is the systematic use of data, statistical methods, computational techniques, and domain knowledge to produce evidence for decisions. Descriptive analytics asks what happened. Diagnostic analytics investigates why it happened. Predictive analytics estimates what is likely to happen. Prescriptive analytics evaluates what should be done, often under operational constraints.

Artificial intelligence refers broadly to computational systems capable of tasks associated with human intelligence, including perception, language processing, reasoning, learning, planning, and generation. Machine learning is a major approach within AI in which models infer patterns from data rather than being programmed entirely through explicit rules. Deep learning uses multilayer neural networks, while generative AI produces new content based on learned representations.

The central concept linking these domains is decision quality. Decision quality is not equivalent to prediction accuracy. It depends on whether the decision objective is correctly specified, whether evidence is relevant and reliable, whether uncertainty is represented honestly, whether alternatives and constraints are considered, and whether consequences are acceptable to affected stakeholders.

4. Fundamental Principles

4.1 Rational choice and bounded rationality

Classical decision theory assumes that a decision maker can compare alternatives consistently and select the action maximizing expected utility. In real organizations, information, time, computational resources, and attention are limited. Herbert Simon's concept of bounded rationality explains why organizations often satisfice: they seek an adequate solution rather than an ideal one. AI can expand analytical capacity, but it does not eliminate institutional constraints or ambiguity about objectives.

4.2 Uncertainty must be represented

Uncertainty may arise from randomness, incomplete information, measurement error, model limitations, or disagreement about the correct model. Treating uncertain estimates as precise facts encourages overconfidence. Probabilities, confidence intervals, prediction intervals, sensitivity analysis, and scenario analysis provide different ways to communicate uncertainty.

4.3 Evidence is conditional on assumptions

Every analysis depends on assumptions about data quality, sampling, measurement, independence, stationarity, missingness, and the relationship between variables. A model is not a neutral window onto reality. Its conclusions are conditional on how the problem was framed and how the data were generated.

4.4 Prediction and causation are distinct

A variable may be highly useful for predicting an outcome without being a cause of it. Causal questions ask what would happen under an intervention: for example, whether increasing training would improve employee performance. Predictive questions ask what outcome is likely for a given case. Confusing the two can produce ineffective or harmful interventions.

4.5 Utility includes more than financial value

Utility may include revenue, cost, safety, equity, privacy, reputation, environmental impact, autonomy, and long-term resilience. Multi-criteria decision analysis makes such dimensions explicit. Weighting them is not merely a technical exercise; it reflects organizational values and should be subject to governance.

4.6 Human judgment is both a resource and a risk

Human experts provide contextual knowledge, ethical interpretation, and responsibility. They may also introduce bias, inconsistency, automation aversion, or automation bias. Effective human–AI collaboration allocates tasks according to comparative strengths and establishes procedures for disagreement, escalation, and accountability.

5. Major Components / Elements

A decision-analysis system normally contains the following elements.

Decision maker and stakeholders. The decision maker may be an individual, team, public agency, or automated service. Stakeholders include people affected by outcomes, even when they do not control the decision.

Objectives and utility function. Objectives define what success means. A utility function maps consequences to preference values. In practice, utilities may be represented through scores, monetary values, thresholds, or multiple criteria.

Actions and feasible alternatives. The set of possible actions must include realistic alternatives and account for legal, technical, financial, and operational constraints.

States of nature. These are uncertain conditions outside direct control, such as demand, disease prevalence, cyberattack, or economic conditions.

Data and evidence. Evidence includes observations, experiments, expert judgments, domain knowledge, and model outputs. Its provenance, quality, relevance, and timeliness must be assessed.

Model. A model represents relationships among inputs, states, actions, and outcomes. It may be statistical, mechanistic, optimization-based, machine-learning-based, or hybrid.

Decision rule. A rule translates evidence into action, such as selecting the alternative with maximum expected utility or referring cases above a risk threshold for human review.

Feedback and monitoring. Decisions change the environment and generate new data. Monitoring is therefore required to detect drift, unintended effects, and changes in stakeholder needs.

6. Process / Framework / Methodology

A rigorous decision and AI methodology can be organized into ten stages.

First, define the decision problem rather than beginning with available data. Specify who must decide, when the decision occurs, what actions are possible, and what consequences matter. A poorly framed question cannot be repaired by a sophisticated model.

Second, identify stakeholders, constraints, and values. Determine legal duties, privacy requirements, safety thresholds, fairness concerns, and acceptable trade-offs. Where values conflict, make the conflict explicit.

Third, construct a causal and operational view of the problem. Identify potential drivers, outcomes, interventions, confounders, and feedback loops. A causal diagram can clarify which variables should be measured and which should not be used as proxies for protected attributes.

Fourth, acquire and govern data. Establish provenance, access controls, documentation, quality checks, retention rules, and procedures for handling missing or sensitive data. Data preparation should be reproducible and auditable.

Fifth, formulate a baseline. A simple rule, historical average, linear model, or expert process provides a comparison. A complex model is valuable only if it improves the decision relative to a credible baseline.

Sixth, select and train models. Model selection should reflect the decision objective, data-generating process, interpretability needs, cost of errors, and deployment environment. Separate training, validation, and test data, and prevent leakage from future or outcome-derived information.

Seventh, evaluate technical and decision performance. Use suitable measures such as calibration, precision, recall, expected loss, subgroup performance, cost-sensitive error, robustness, and operational impact. Accuracy alone is rarely sufficient.

Eighth, conduct sensitivity, scenario, and stress analysis. Examine how recommendations change under alternative assumptions, data perturbations, distribution shifts, and extreme conditions.

Ninth, deploy with controls. Define human roles, override procedures, access permissions, logging, incident response, and communication to affected users. Deployment is a socio-technical intervention, not merely a software release.

Tenth, monitor, learn, and govern. Track performance, drift, fairness, security, user behavior, and downstream outcomes. Establish conditions for retraining, suspension, independent review, and retirement.

7. Important Theories and Models

7.1 Expected utility theory

If action a can produce consequences under states s, expected utility is commonly expressed as:

EU(a) = Σ P(s | E) U(a,s)

where P(s | E) is the probability of state s given evidence E, and U(a,s) is the utility of taking action a when state s occurs. The preferred action maximizes expected utility, subject to constraints. This framework incorporates risk preferences. A risk-averse organization may prefer a lower but more reliable outcome to a higher expected value with substantial downside.

7.2 Bayesian reasoning

Bayesian reasoning updates prior beliefs using evidence. In simplified form:

P(H | E) = [P(E | H)P(H)] / P(E)

A prior probability represents knowledge before new evidence; the likelihood indicates how compatible the evidence is with a hypothesis; the posterior represents the updated belief. Bayesian reasoning is useful in diagnosis, fraud detection, forecasting, and scientific inference. Its value lies not in producing certainty but in requiring explicit treatment of base rates and evidence strength.

7.3 Value of information

Information has value when it can change an action and improve expected utility. The expected value of perfect information is the difference between the expected utility with perfect knowledge and the maximum expected utility without it. Expected value of sample information considers the likely quality and cost of an additional study or test. This helps managers decide whether further data collection is worth its cost.

7.4 Optimization

Optimization identifies the best feasible solution according to an objective function. Linear programming, integer programming, nonlinear optimization, dynamic programming, and robust optimization are used in scheduling, logistics, portfolio management, network design, and resource allocation. Optimization requires explicit objectives and constraints; it cannot determine what should be valued unless those values are supplied by decision makers.

7.5 Statistical inference and machine learning

Statistical inference often emphasizes estimation, uncertainty, explanation, and generalization from samples to populations. Machine learning often emphasizes predictive performance on future observations. The distinction is not absolute: modern statistical learning combines both perspectives. Cross-validation, regularization, ensembles, and representation learning improve prediction, while hierarchical models, causal models, and uncertainty quantification support interpretation and inference.

7.6 Causal inference

Causal inference estimates the effect of an intervention or treatment. Randomized experiments provide strong identification when properly designed. Observational methods include matching, regression adjustment, instrumental variables, regression discontinuity, difference-in-differences, and causal graphical models. The validity of each method depends on assumptions such as exchangeability, consistency, positivity, and correct temporal ordering.

7.7 Explainability and interpretability

Interpretability concerns how directly a model can be understood, while explainability concerns methods used to communicate or approximate its reasoning. Intrinsically interpretable models include decision trees, generalized additive models, and sparse linear models. Post-hoc tools include feature attribution, partial dependence, counterfactual explanations, and surrogate models. Explanations should be evaluated for fidelity, stability, comprehensibility, and usefulness rather than accepted merely because they are visually appealing.

7.8 Fairness and calibration

Fairness is multidimensional. Possible criteria include equal error rates, demographic parity, equal opportunity, individual fairness, and calibration. In many settings these criteria cannot all be satisfied simultaneously when groups differ in base rates. Therefore, fairness requires a normative decision about which harms matter and how they should be balanced, supported by empirical monitoring and stakeholder participation.

8. Practical Applications

In healthcare, analytics can support diagnosis, patient-risk stratification, staffing, and treatment planning. The appropriate target may be calibrated risk rather than raw classification accuracy, and causal evidence is needed before recommending an intervention.

In finance, models estimate credit risk, detect fraud, forecast liquidity, and optimize portfolios. Governance must address explainability, discrimination, model risk, market instability, and regulatory reporting.

In manufacturing, predictive maintenance models estimate equipment failure, while optimization systems schedule maintenance under production constraints. A false alarm may be costly, but a missed failure may threaten safety and continuity.

In public administration, AI may assist benefit allocation, tax compliance, emergency response, and policy evaluation. Public-sector use requires heightened transparency, procedural fairness, contestability, and protection against unjustified automation.

In cybersecurity, Bayesian updating, anomaly detection, threat intelligence, and optimization support prioritization of incidents. Since adversaries adapt, models must be stress-tested against strategic manipulation and distribution shift.

In education and human resources, models may identify support needs or forecast attrition. Such applications require caution because historical outcomes may reflect unequal opportunities rather than inherent characteristics.

9. Examples

Consider a hospital deciding whether to admit a patient for observation. A predictive model estimates a 12 percent probability of deterioration. The decision cannot be made from this probability alone. The hospital must consider the cost of admission, the cost of missed deterioration, bed availability, patient preferences, and the reliability of the model for relevant subgroups. A Bayesian approach may update risk with new vital signs; a decision rule may refer high-risk cases to clinicians; and a causal analysis may be required before changing treatment protocols.

Consider an online retailer choosing delivery capacity. Demand forecasts provide probability distributions rather than a single number. An optimization model can allocate vehicles and warehouses under capacity constraints. The utility function may include delivery speed, cost, emissions, and customer retention. Scenario analysis can test disruptions such as weather or fuel-price shocks.

Consider an employer using a model to identify employees likely to leave. The model may predict attrition accurately but could encourage intrusive surveillance or discriminatory intervention. A causal evaluation is needed to determine whether proposed retention actions actually improve retention, rather than merely identifying people who were already likely to leave.

10. Technology and Digital Applications

Relevant technologies include data warehouses and lakehouses for governed analytical data, statistical computing environments, machine-learning platforms, optimization solvers, causal-inference libraries, model registries, and monitoring systems. MLOps practices connect model development with reproducible deployment, version control, testing, observability, and controlled retraining.

Generative AI systems based on large language or multimodal models can summarize documents, draft code, support research exploration, and provide conversational interfaces to enterprise knowledge. Retrieval-augmented generation can connect outputs to approved sources, although retrieval does not guarantee correct interpretation. Guardrails, access control, prompt and output logging, citation verification, and human review are important in high-impact contexts.

Digital twins and simulation environments allow organizations to examine policies and operational changes before implementation. Edge AI supports low-latency decisions near sensors or devices, while privacy-enhancing technologies such as federated learning, secure computation, and differential privacy can reduce exposure of sensitive data. These technologies remain subject to trade-offs among utility, cost, latency, privacy, and governance.

11. Strategic / Managerial Importance

Decision intelligence can improve resource allocation, responsiveness, service quality, and organizational learning. It enables managers to move from retrospective reporting toward scenario planning and prescriptive action. However, value arises only when analytical outputs are integrated into real workflows and linked to measurable outcomes.

Managers must establish an AI portfolio rather than treating each model as an isolated experiment. Portfolio governance should prioritize use cases by value, feasibility, risk, data readiness, and strategic alignment. Leaders must also invest in data stewardship, interdisciplinary talent, process redesign, cybersecurity, and organizational literacy. A model that staff do not trust, understand, or know how to use will not create sustained value.

Strategically, organizations should preserve optionality. They should avoid dependence on a single vendor or opaque model, maintain contingency procedures, and consider how competitors, regulators, employees, and customers may respond. Responsible adoption can become a source of trust and competitive differentiation.

12. Detailed Case Study

The following case is fictional and illustrative.

Background

Northbridge Transit Authority operates buses and light rail across a large metropolitan region. Service interruptions are expensive, and unplanned vehicle failures create safety and reputational risks. The authority possesses maintenance records, sensor data, weather information, route schedules, and technician reports.

Problem

The authority proposed a predictive-maintenance system that would identify vehicles likely to fail within fourteen days. The initial vendor claimed 92 percent accuracy. Managers needed to determine whether the system should be adopted, how it should influence maintenance schedules, and how to prevent unequal service impacts across routes.

Analysis

The project team first discovered that the claimed accuracy was calculated on randomly split historical data. Because multiple records from the same vehicle appeared in both training and test sets, the estimate suffered from leakage. A time-based evaluation reduced performance substantially. The team also found that vehicles assigned to lower-income routes had more frequent sensor outages, producing less reliable predictions.

A causal review showed that maintenance history was partly a consequence of prior managerial decisions. Using it without adjustment could reproduce historical allocation patterns. The team therefore separated prediction from intervention: the model estimated failure risk, while a constrained optimization model scheduled inspections using risk, safety criticality, route coverage, spare capacity, and equity requirements.

The authority compared the model with a simple age-and-mileage rule. It evaluated calibration, false negatives, false positives, subgroup performance, downtime, maintenance cost, and missed-service hours. A pilot included human review by experienced technicians. Technicians could override recommendations but had to record reasons, creating evidence for later evaluation.

Strategy / Solution

Northbridge adopted a staged strategy. The model was used first as a decision-support tool rather than an autonomous controller. Sensor-quality indicators were displayed alongside risk scores. Vehicles with poor data quality were not treated as low risk; they were referred for inspection or sensor repair. The optimization system imposed minimum service coverage and prohibited deferring safety-critical maintenance solely for efficiency.

The authority established a model-risk committee, quarterly performance reviews, incident reporting, and an annual independent audit. It required the vendor to provide documentation, data lineage, update procedures, and access to evaluation results. A public summary explained the purpose of the system without exposing security-sensitive details.

Outcome

After six months, unplanned failures declined by 18 percent and maintenance became more evenly distributed across depots. The authority did not achieve the vendor's original accuracy claim, but it achieved a measurable reduction in disruption without increasing missed-service hours on lower-income routes. Several false alarms led to sensor repairs that improved future data quality. The pilot also showed that technicians were more willing to use the system when explanations identified relevant signals and displayed uncertainty.

Learning Points

The case demonstrates that decision value is different from headline model accuracy. Temporal validation, baseline comparison, causal awareness, fairness monitoring, optimization constraints, human oversight, and continuous evaluation were all necessary. It also illustrates that data quality is itself a decision variable: unreliable measurements should trigger information-gathering or maintenance actions rather than confident predictions.

13. Advantages / Benefits

Integrated decision theory and AI can improve consistency, speed, scalability, forecasting, resource utilization, early warning, and scenario analysis. Models can reveal patterns that exceed unaided human attention and can support decisions across geographically distributed operations. Bayesian and probabilistic methods make uncertainty more visible, while optimization can coordinate complex constraints. Properly designed human–AI systems can combine computational breadth with human contextual judgment.

14. Limitations / Challenges

Data may be incomplete, biased, stale, or generated by processes that change when a model is introduced. Models can overfit, drift, fail under distribution shift, or perform unevenly across groups. Proxy variables may reproduce protected characteristics. Optimization can produce technically optimal but socially unacceptable outcomes when objectives are incomplete. Generative AI may hallucinate, expose confidential information, reproduce training biases, or generate plausible but unsupported reasoning.

Additional challenges include cybersecurity attacks, adversarial manipulation, vendor opacity, high computational cost, unclear liability, automation bias, deskilling, and difficulty measuring long-term effects. Explainability methods may be unstable or misleading, and fairness criteria may conflict. Finally, organizations may lack the governance capacity required to monitor systems after deployment.

15. Ethical and Governance Considerations

Ethical AI requires respect for autonomy, privacy, safety, fairness, human dignity, and accountability. Organizations should establish purpose limitation, data minimization, access control, retention policies, and meaningful consent where appropriate. High-impact decisions should offer notice, explanation proportionate to the context, review, and avenues for contesting an outcome.

Governance should assign responsibility across developers, data owners, managers, users, vendors, and senior leadership. Documentation should cover intended use, prohibited use, data sources, assumptions, evaluation results, limitations, and change history. Independent review is especially important when systems affect employment, credit, healthcare, education, policing, or public benefits.

Security governance must address prompt injection, data poisoning, model extraction, unauthorized access, and supply-chain risk. Generative AI requires controls against confidential-data leakage, fabricated citations, copyright violations, and unapproved autonomous action. Ethical review should continue after deployment because real-world effects may differ from laboratory performance.

16. Current and Emerging Trends

Important trends include foundation models, multimodal AI, retrieval-augmented generation, autonomous agents, synthetic data, causal machine learning, federated learning, edge intelligence, and decision platforms integrating prediction with optimization. Digital twins increasingly combine real-time data, simulation, and operational control.

There is growing emphasis on evaluation beyond benchmark accuracy, including robustness, calibration, red-teaming, energy use, subgroup effects, and real-world utility. Regulatory developments are encouraging risk classification, documentation, transparency, human oversight, and accountability. Organizations are also exploring smaller specialized models that may offer lower cost, better privacy, and greater domain control than general-purpose systems.

17. Best Practices

Begin with a consequential decision and a clearly defined objective, not with enthusiasm for a particular technology. Establish a credible baseline and specify the costs of different errors. Use representative, governed data and document its provenance. Separate training, validation, and temporal testing data. Evaluate calibration, subgroup performance, robustness, uncertainty, and actual decision outcomes.

Use causal methods when the question concerns intervention. Combine prediction with optimization only when objectives and constraints are explicit. Prefer interpretable models where performance is adequate, and test post-hoc explanations for fidelity. Design human oversight with defined authority, training, workload limits, and escalation procedures. Monitor for drift, unintended impacts, security incidents, and changes in context. Permit independent audit, user feedback, appeal, model suspension, and orderly retirement.

18. Future Scope

Future research will increasingly integrate causal reasoning, learning, simulation, and optimization into adaptive decision systems. These systems may estimate not only what is likely to happen but also which intervention is most effective under changing conditions. Advances in uncertainty-aware foundation models could improve reliability, although fundamental limitations in data and reasoning will remain.

Research is also needed on collective human–AI intelligence, institutional accountability, algorithmic market effects, energy-efficient learning, privacy-preserving analytics, and democratic governance. The central future challenge is not simply making AI more capable. It is designing socio-technical institutions in which capability is aligned with evidence, human values, resilience, and legitimate authority.

19. Chapter Summary

Decision theory provides principles for selecting actions under uncertainty; analytics converts data into descriptive, predictive, and prescriptive insight; and AI supplies computational methods for learning, reasoning, and generation. Their integration requires explicit objectives, utility, probabilities, constraints, causal reasoning, and continuous evaluation. Prediction is not causation, accuracy is not decision value, and explainability is not automatically accountability. Effective systems combine technical validation with ethical governance, human judgment, stakeholder participation, and operational monitoring. Responsible AI is therefore a discipline of evidence-based decision design rather than an algorithm-selection exercise.

20. Key Terms

1. Decision theory: The study of rational action under certainty, risk, and uncertainty.

2. Utility: A representation of the value or preference associated with an outcome.

3. Expected utility: The probability-weighted utility of an action across possible states.

4. Bayesian reasoning: Updating beliefs by combining prior knowledge with evidence.

5. Value of information: The improvement in expected decision value obtained from additional information.

6. Analytics: The systematic use of data and methods to support understanding and decisions.

7. Predictive analytics: Estimation of likely future or unknown outcomes.

8. Prescriptive analytics: Analysis that recommends actions under objectives and constraints.

9. Optimization: Selection of the best feasible solution according to a defined objective.

10. Machine learning: Methods that learn patterns or functions from data.

11. Causal inference: Estimation of effects produced by interventions or treatments.

12. Calibration: Agreement between predicted probabilities and observed frequencies.

13. Model drift: Degradation or change in model behavior as data or context changes.

14. Explainability: The capacity to communicate or approximate how a model produces outputs.

15. Interpretability: The degree to which a model's operation can be understood directly.

16. Generative AI: AI that creates new content such as text, images, audio, or code.

17. Human–AI collaboration: Coordinated performance in which humans and AI contribute complementary capabilities.

18. Distribution shift: A change between the data used for development and the data encountered in operation.

21. Review Questions

Short Answer Questions

1. What is the difference between predictive and prescriptive analytics?

2. Why is a baseline important when evaluating an AI model?

3. Define expected utility and identify its main components.

4. Why does correlation not establish causation?

5. What is model drift, and why should it be monitored?

Long Answer Questions

1. Explain how Bayesian reasoning and value-of-information analysis can improve managerial decisions under uncertainty.

2. Compare statistical inference, machine learning, and causal inference in terms of purpose, assumptions, and evaluation.

3. Discuss why fairness criteria may conflict and explain how an organization should govern such trade-offs.

4. Analyze the role of optimization in transforming predictions into operational decisions.

5. Evaluate the risks and opportunities of generative AI in a university or enterprise setting.

Analytical / Application Questions

1. A bank's credit model has high overall accuracy but substantially higher false-negative rates for one demographic group. Propose an evaluation and governance process that addresses performance, fairness, explainability, and business constraints.

2. A public health agency must decide whether to launch a costly prevention program based on uncertain disease forecasts. Construct a decision framework using probabilities, utilities, causal evidence, and value-of-information analysis.

3. Design a human–AI workflow for predictive maintenance in a manufacturing firm. Specify the data, baseline, model evaluation, optimization constraints, human roles, monitoring indicators, and conditions for suspending the system.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes prediction from causal inference?

A. Prediction estimates likely outcomes, whereas causal inference estimates effects of interventions.

B. Prediction always requires randomized experiments, whereas causal inference never uses data.

C. Prediction describes utility, whereas causal inference calculates financial profit.

D. Prediction is qualitative, whereas causal inference is exclusively algorithmic.

Correct Answer: A. Prediction estimates likely outcomes, whereas causal inference estimates effects of interventions.

Question 2

A manager uses prior disease prevalence together with a new diagnostic test result to revise a patient's risk estimate. Which principle is being applied?

A. Linear programming

B. Bayesian reasoning

C. Demographic parity

D. Data minimization

Correct Answer: B. Bayesian reasoning

Question 3

Why is a credible baseline important in machine-learning evaluation?

A. It guarantees that the model is unbiased.

B. It eliminates the need for validation data.

C. It shows whether model complexity produces meaningful improvement over a simpler approach.

D. It converts observational data into experimental data.

Correct Answer: C. It shows whether model complexity produces meaningful improvement over a simpler approach.

Question 4

What does calibration measure in a probabilistic prediction model?

A. Whether predicted probabilities correspond to observed frequencies

B. Whether the model has the fewest possible input variables

C. Whether all demographic groups have identical base rates

D. Whether a model can generate synthetic data

Correct Answer: A. Whether predicted probabilities correspond to observed frequencies

Question 5

Which action is most appropriate when sensor data are unreliable in a predictive-maintenance system?

A. Treat every unreliable record as low risk

B. Remove all affected vehicles from evaluation permanently

C. Trigger data-quality review or inspection rather than making a confident risk judgment

D. Increase the model's confidence score automatically

Correct Answer: C. Trigger data-quality review or inspection rather than making a confident risk judgment

Question 6

What is the principal purpose of optimization after a risk model has produced predictions?

A. To replace all organizational objectives with accuracy

B. To select feasible actions while considering objectives and constraints

C. To prove that predictions are causal

D. To guarantee equal outcomes for every stakeholder

Correct Answer: B. To select feasible actions while considering objectives and constraints

Question 7

Which issue is especially associated with generative AI systems?

A. They cannot process any unstructured information.

B. They may produce fluent but fabricated or unsupported content.

C. They always provide transparent reasoning.

D. They eliminate the need for human accountability.

Correct Answer: B. They may produce fluent but fabricated or unsupported content.

Question 8

Why can a highly accurate predictive model still be inappropriate for a high-impact decision?

A. Accuracy is unrelated to all decisions.

B. The model may be unfair, poorly calibrated, non-causal for the intended intervention, or operationally unsafe.

C. High accuracy proves that the data are invalid.

D. High-impact decisions cannot use probabilities.

Correct Answer: B. The model may be unfair, poorly calibrated, non-causal for the intended intervention, or operationally unsafe.

Question 9

Which governance practice best supports responsible deployment of an AI system?

A. Deploy once and avoid changing the model

B. Monitor drift, subgroup outcomes, security, and real-world decision effects

C. Hide model limitations to preserve user confidence

D. Allow the vendor to define all organizational objectives

Correct Answer: B. Monitor drift, subgroup outcomes, security, and real-world decision effects

Question 10

What is the expected value of information concerned with?

A. The cost of storing historical data

B. The improvement in expected decision value that additional information may provide

C. The number of parameters in a neural network

D. The probability that a model is explainable

Correct Answer: B. The improvement in expected decision value that additional information may provide