1. Foundations of AI Ethics and Responsible AI

Introduction and Context

Artificial intelligence is no longer confined to research laboratories. It influences credit decisions, recruitment, medical diagnosis, insurance pricing, education, policing, content recommendation, industrial operations, and public administration. These applications can improve efficiency and extend human capabilities, but they can also distribute opportunities and risks unevenly. An inaccurate medical model may delay treatment; a biased recruitment system may exclude qualified applicants; an opaque public-sector algorithm may affect access to essential services without providing a meaningful explanation.

The central question of AI ethics is therefore not simply whether a system can be built, but whether it should be built, how it should be used, who may be affected, and who is answerable when harm occurs. Responsible AI translates ethical reflection into organizational practices, technical controls, governance arrangements, and professional duties. It treats AI as a socio-technical system: a system whose outcomes arise from interactions among data, models, software, institutions, users, policies, incentives, and wider social conditions.

This chapter establishes the conceptual foundation for the study of AI ethics, governance, and explainable AI. It distinguishes ethical, legal, technical, and managerial responsibilities; introduces major principles and models; and shows how accountability must operate throughout the AI lifecycle.

Learning Objectives

By the end of this chapter, learners should be able to:

1. Define AI ethics, responsible AI, trustworthy AI, algorithmic accountability, and explainability.

2. Explain why AI must be understood as a socio-technical system.

3. Distinguish ethical, legal, technical, and managerial responsibilities in AI development and use.

4. Apply major responsible-AI principles to practical situations.

5. Identify how accountability can be implemented across the AI lifecycle.

6. Analyse the risks, limitations, and trade-offs associated with responsible AI.

7. Evaluate organizational practices for developing and deploying trustworthy AI systems.

Why AI Ethics Is a Systems Question

An AI model is often described as a mathematical function that maps inputs to outputs. This description is useful but incomplete. The practical effect of a model depends on what data were collected, how labels were created, which objective was selected, how thresholds were set, who interprets the result, and what institutional action follows it.

Consider a system designed to identify potentially fraudulent insurance claims. The model may achieve high overall accuracy, yet still produce excessive false accusations for a particular community if historical investigations were concentrated in that community. The problem may not be reducible to code. It may involve biased historical practices, an incomplete definition of fraud, poor data governance, pressure to reduce costs, and a workflow in which employees treat a risk score as a final decision. Correcting the algorithm alone may not correct the institutional process.

This is the significance of the socio-technical perspective. It directs attention to at least five connected elements:

• Technical artefacts: data pipelines, models, interfaces, infrastructure, and monitoring tools.
• Human actors: developers, domain experts, decision-makers, affected individuals, and oversight personnel.
• Organizational arrangements: objectives, incentives, procurement procedures, reporting lines, and risk controls.
• Social and institutional context: historical inequalities, professional norms, public expectations, and power relationships.
• Legal and regulatory environment: rights, duties, sectoral standards, and remedies.

An ethical assessment that examines only model performance is consequently inadequate. A technically accurate system can still be unjust, unsafe, unlawful, or inappropriate for its intended use.

Core Concepts and Their Relationships

AI Ethics

AI ethics is the systematic study and practical application of moral principles to the design, development, deployment, use, and retirement of AI systems. It asks questions about harm, fairness, autonomy, dignity, responsibility, privacy, power, and the distribution of benefits and burdens.

AI ethics is broader than compliance. Law establishes enforceable minimum requirements, while ethics may demand more responsible conduct even where a practice is technically lawful. For example, collecting extensive personal data may be legally permissible under a particular arrangement but ethically questionable if individuals cannot reasonably understand or refuse the practice.

Responsible AI

Responsible AI refers to the policies, processes, technical methods, and organizational behaviours used to ensure that AI systems are developed and operated in accordance with ethical values, legal duties, professional standards, and legitimate societal expectations. It is an action-oriented concept. It includes impact assessments, documentation, testing, human oversight, incident response, stakeholder engagement, and continuous monitoring.

Responsible AI is not a single software feature. It is a lifecycle discipline. A model cannot be made responsible merely by adding an explanation screen after deployment if its purpose, data, or institutional use is fundamentally inappropriate.

Trustworthy AI

Trustworthy AI is AI that merits justified confidence because it operates reliably, safely, transparently enough for its context, fairly, securely, and under accountable human and organizational control. Trust is not identical to user comfort or popularity. A system may be trusted by users for the wrong reasons, while a trustworthy system may appropriately communicate uncertainty and limitations.

Trustworthiness is contextual. The level of robustness, explainability, human review, and evidence required for an automated playlist recommendation differs from that required for an immigration, employment, or clinical decision-support system.

Algorithmic Accountability

Algorithmic accountability is the obligation and capacity of identifiable people and institutions to explain, justify, monitor, and take responsibility for the consequences of algorithmic systems. It includes answerability—being required to provide reasons and evidence—and enforceability—the existence of consequences, corrective action, or remedies when obligations are breached.

Accountability should not be assigned vaguely to “the algorithm.” Algorithms do not independently bear legal or moral responsibility. Responsibility rests with the organizations and professionals that commission, build, configure, deploy, supervise, and rely on them. Accountability can nevertheless be distributed across several actors, provided that roles are clear and do not create gaps in responsibility.

Explainability and Interpretability

Explainability concerns the extent to which the behaviour, output, or reasoning of an AI system can be meaningfully communicated to relevant people. Interpretability often refers more narrowly to the extent to which a model’s operation can be understood directly, such as through a small decision tree or an understandable linear model.

An explanation may describe:

• Function: what the system is designed to do.
• Data: what information was used and how it was obtained.
• Logic: how relevant factors influenced an output.
• Outcome: why a particular decision or recommendation was produced.
• Process: how a person can challenge, correct, or appeal the result.

An explanation is not automatically a justification. A technically accurate account of a discriminatory process does not make that process acceptable. Nor is more explanation always better: excessive technical detail may confuse affected individuals and conceal the practically important information.

Ethical Foundations for Responsible AI

Different frameworks use different labels, but several ethical concerns recur across professional and policy discussions.

Respect for autonomy requires that people not be manipulated, deceived, or subjected to consequential automated decisions without appropriate information and meaningful choice where choice is feasible. Autonomy also involves preserving the ability to contest or decline inappropriate uses of automation.

Beneficence and non-maleficence require systems to produce genuine benefits and avoid or reduce foreseeable harm. Harm may be physical, psychological, economic, reputational, social, environmental, or political. Safety testing must therefore consider more than system failure in a narrow engineering sense.

Justice and fairness concern the distribution of benefits, burdens, opportunities, and errors. Fairness cannot always be represented by one metric. Equal treatment, equal outcomes, equal error rates, individual consistency, and protection against historical disadvantage may conflict. Selecting a fairness criterion is a normative and contextual decision, not merely a statistical one.

Respect for dignity requires treating people as more than data points or optimization targets. Systems that classify individuals may undermine dignity when they reduce complex identities to stigmatizing categories or deny people a voice in decisions affecting their lives.

Privacy involves appropriate collection, use, disclosure, retention, and inference from personal information. Privacy is not limited to secrecy. It includes contextual integrity, control over personal information, and protection against intrusive profiling.

Solidarity and inclusion draw attention to people who may be excluded by language, disability, poverty, limited digital access, or inadequate representation in data. Systems should be assessed for accessibility and for effects on communities, not only on direct users.

These principles are not a checklist that automatically resolves every conflict. A fraud-control system may improve collective security while imposing additional burdens on innocent customers. Ethical analysis must make such trade-offs visible, assess their proportionality, and provide safeguards.

A Lifecycle Model of Responsible AI

Responsible AI is best implemented through decisions made at successive but interconnected stages.

Purpose and Problem Definition

Before selecting a model, an organization should establish the legitimate purpose of the system, the decision it will support, the people affected, and the alternatives available. A useful question is whether automation is necessary at all. A technically feasible system may still be socially unnecessary or disproportionate.

Data and Representation

Teams should examine data provenance, consent or lawful basis, representativeness, quality, missingness, historical bias, proxy variables, and labeling practices. Data documentation should record known limitations rather than presenting a dataset as neutral.

Model Design and Development

The choice of model, target variable, optimization objective, and threshold involves value judgments. Developers should consider performance across relevant groups, uncertainty, robustness, privacy, security, and the consequences of false positives and false negatives. Simpler models may be preferable where they provide sufficient performance and greater interpretability.

Validation and Impact Assessment

Testing should combine quantitative evaluation with domain expertise and stakeholder perspectives. Impact assessment asks who may be harmed, how severe the harm could be, whether benefits justify risks, and what mitigations are available. Independent review is particularly important for high-impact applications.

Deployment and Human Use

Deployment includes interfaces, workflows, training, access controls, escalation procedures, and communication with affected people. Human oversight is meaningful only when reviewers have authority, competence, time, and access to relevant information. A nominal human-in-the-loop who merely approves every automated output does not provide effective oversight.

Monitoring, Incident Response, and Retirement

Performance and impacts may change after deployment because populations, behaviour, policies, or data distributions change. Organizations should monitor drift, disparate error patterns, complaints, security events, and unexpected uses. There must be procedures for correcting data, suspending a system, notifying affected parties, and providing remedies. Retirement is also a governance decision: systems should not remain in operation simply because they already exist.

Distinguishing Responsibilities

Ethical Responsibility

Ethical responsibility concerns what ought to be done in light of values such as fairness, dignity, welfare, autonomy, and justice. It may apply even when no law directly prohibits an action. A product manager who recognizes that a recommendation system promotes harmful content has an ethical responsibility to investigate and mitigate the harm, even before regulation or litigation occurs.

Legal and Regulatory Responsibility

Legal responsibility arises from statutes, regulations, contracts, case law, and recognized rights. It may involve data protection, discrimination, consumer protection, intellectual property, sectoral safety, employment law, or administrative law. Compliance is necessary but not sufficient. Laws can be incomplete, jurisdiction-specific, or slower than technological change.

Technical Responsibility

Technical responsibility includes building systems that are secure, robust, tested, documented, maintainable, and appropriate for their intended context. It includes selecting suitable metrics, controlling access, validating data pipelines, evaluating model performance, and communicating limitations. Technical competence does not eliminate ethical responsibility, but poor engineering can turn ethical commitments into empty statements.

Managerial and Organizational Responsibility

Managers determine objectives, budgets, deadlines, staffing, procurement, risk tolerance, and escalation routes. They are responsible for creating conditions in which ethical concerns can be raised without retaliation. Organizations must assign ownership, maintain records, conduct reviews, and ensure that commercial pressure does not silently override safety or rights.

These responsibilities overlap but cannot substitute for one another. A developer cannot transfer all responsibility to a manager, and a legal team cannot guarantee fairness merely by approving documentation. Effective governance connects professional roles through explicit decision rights and escalation mechanisms.

Governance Structures and Practical Tools

AI governance is the institutional framework through which an organization directs and controls AI-related activities. It should be proportionate to risk and integrated with existing risk, security, privacy, compliance, and quality-management systems.

A practical governance arrangement may include an accountable executive, a cross-functional review committee, system owners, technical leads, domain specialists, privacy and security officers, and channels for affected individuals to challenge decisions. Clear records should identify the intended purpose, model version, data sources, evaluation results, approvals, incidents, and changes after deployment.

Useful tools include:

• Algorithmic impact assessments, which examine likely effects on rights, groups, safety, and access.
• Datasheets or data statements, which document data origin, composition, collection conditions, intended uses, and limitations.
• Model cards or system documentation, which describe performance, evaluation populations, appropriate uses, failure modes, and restrictions.
• Bias and fairness testing, using contextually justified measures rather than relying on a single score.
• Red-team exercises, which deliberately search for harmful, insecure, discriminatory, or deceptive behaviour.
• Audit logs and version control, which support reconstruction of how outputs were produced.
• Appeal and remediation mechanisms, which allow correction, human review, and compensation or other remedies where appropriate.

Governance should be risk-based. A low-impact movie recommender and an automated benefits-eligibility system should not face identical approval requirements. Risk classification should consider severity, scale, reversibility, vulnerability of affected people, degree of automation, and possibility of meaningful human correction.

Practical Applications

In recruitment, responsible AI requires more than removing explicit gender or racial fields. Other variables may act as proxies, historical labels may encode biased hiring patterns, and the interface may encourage recruiters to over-rely on rankings. Appropriate practice includes job-related feature selection, subgroup evaluation, accessibility testing, human review, candidate communication, and a route for correction.

In healthcare, an AI system should be evaluated for clinical validity, subgroup performance, data shift across hospitals, privacy, workflow integration, and the consequences of erroneous recommendations. Clinicians need to understand when not to rely on the system. A model that performs well in a research dataset may not be suitable for a different population or clinical setting.

In financial services, explainability and contestability are particularly important because automated credit or fraud decisions affect economic opportunity. Individuals may need a specific, actionable reason for an adverse decision rather than a generic statement that the model produced a low score.

In public administration, governance must address due process, transparency, equality before public institutions, procurement accountability, and the ability of citizens to challenge decisions. Public agencies should not treat a vendor’s claim of trade secrecy as a complete justification for denying meaningful oversight.

Illustrative Case Study: Automated Benefits Eligibility Screening

Background

A fictional regional government introduces an AI-assisted system to prioritize applications for housing assistance. The system estimates the likelihood that an application requires urgent review and allocates cases to different processing queues.

Situation and Problem

After deployment, average processing time decreases. However, community organizations report that applicants with unstable housing histories and those who frequently change telephone numbers are disproportionately classified as low priority. The model uses historical case records, contact stability, prior administrative interactions, and geographic information. Staff members increasingly rely on the priority score because of workload pressures.

Analysis

The system’s objective—faster allocation of scarce administrative attention—may be legitimate, but the target and features create risks. Historical administrative records may reflect unequal access to services rather than actual need. Telephone instability may be associated with poverty and therefore function as a proxy for socioeconomic disadvantage. Geographic variables may reproduce segregation. The model also changes the practical meaning of a low-priority classification: it is not merely a prediction but a gateway to delayed assistance.

The problem is therefore socio-technical. It includes data limitations, an organizational incentive to reduce queues, inadequate explanation to applicants, and weak human review. High aggregate accuracy would not resolve the issue.

Approach and Strategy

The agency pauses automatic deprioritization and uses the model only to identify cases for additional review. It conducts an impact assessment with applicants, caseworkers, disability advocates, and privacy specialists. Features are reassessed for necessity and proxy effects. Performance is evaluated separately across relevant groups and under changing contact conditions. Staff are trained to treat scores as uncertain evidence rather than decisions.

The agency introduces a written explanation process, a rapid appeal channel, audit logs, periodic independent review, and a rule that indicators of vulnerability cannot be used to reduce access without documented justification. It also establishes a system owner with authority to suspend the tool when harm indicators exceed predefined thresholds.

Outcome

The revised system continues to assist with workload management but no longer determines priority automatically. Monitoring identifies fewer inappropriate low-priority classifications, while appeals provide information for correcting records and improving the service. The agency also discovers that a non-technical intervention—simplifying application procedures—reduces delays more effectively for some applicants than further model optimization.

Learning Points

The case demonstrates that responsible AI begins with problem formulation, not model selection. It also shows that fairness requires attention to institutional history, that human oversight must have real authority, and that explainability should support contestability. Finally, governance may reveal that redesigning the surrounding process is more beneficial than improving the algorithm alone.

Challenges, Limitations, and Common Mistakes

One common mistake is treating fairness as a single numerical property. Different fairness criteria can conflict, and a model may satisfy a metric while producing unfair outcomes because the target or decision process is inappropriate. Metrics must therefore be interpreted with domain knowledge and stakeholder input.

Another mistake is confusing transparency with accountability. Publishing source code or a general description may not enable an affected person to understand a decision or obtain a remedy. Accountability requires responsible actors, evidence, review, and consequences.

Organizations also frequently rely on “human oversight” symbolically. If reviewers lack time, authority, training, or access to alternatives, they may simply defer to the system. Automation bias can make human involvement less protective than expected.

Other persistent challenges include proprietary systems that limit independent auditing, incomplete or changing data, uncertain causal relationships, privacy risks created by detailed documentation, conflicting international requirements, and pressure to deploy before evaluation is complete. Responsible AI cannot eliminate all uncertainty. Its purpose is to make uncertainty visible, manage it proportionately, and prevent avoidable harm.

Current Developments and Emerging Directions

Current practice is moving from voluntary principles toward operational standards, sector-specific regulation, assurance methods, and documented risk-management processes. Organizations increasingly combine model evaluation with impact assessments, incident reporting, procurement controls, and post-deployment monitoring.

Generative AI has expanded the scope of responsible-AI concerns. In addition to discrimination and privacy, organizations must consider fabricated outputs, intellectual-property issues, prompt manipulation, data leakage, impersonation, and the difficulty of tracing content provenance. Foundation models also raise questions about responsibility across providers, integrators, deployers, and end users.

A significant emerging direction is the development of AI assurance: structured evidence that a system meets stated requirements for safety, security, fairness, privacy, reliability, and explainability. Assurance is most credible when it is continuous, independently challengeable, and connected to real remedies rather than limited to a pre-deployment certificate.

Best Practices for Professionals and Organizations

Professionals should begin with a clear statement of purpose and reject uses that are disproportionate to the social benefit claimed. They should document assumptions, test relevant subgroups, preserve auditability, communicate uncertainty, and seek domain and stakeholder expertise early.

Organizations should assign accountable owners, use risk-based review, maintain lifecycle documentation, establish independent escalation, train staff, monitor real-world outcomes, and provide accessible appeal mechanisms. Procurement contracts should require information about data, performance, security, updates, incidents, audit rights, and responsibility for remediation.

The most mature approach treats responsible AI as part of ordinary organizational quality and risk management rather than as a separate public-relations exercise. Ethical commitments become credible when they influence budgets, release decisions, performance indicators, employee incentives, and the willingness to stop or redesign a system.

Chapter Summary

AI ethics examines the moral questions raised by AI, while responsible AI puts ethical, legal, technical, and organizational commitments into practice across the system lifecycle. Trustworthy AI requires justified confidence based on appropriate performance, safety, fairness, security, transparency, and human control. Algorithmic accountability means that identifiable actors and institutions must explain, justify, monitor, and remedy the effects of AI systems.

AI systems are socio-technical: their outcomes depend on data, models, people, organizations, institutions, and context. Ethical principles such as autonomy, welfare, non-maleficence, justice, dignity, privacy, and inclusion guide analysis but do not mechanically resolve trade-offs. Effective governance begins with problem definition and continues through data management, development, validation, deployment, monitoring, incident response, and retirement. Meaningful explainability must support understanding and contestability, while human oversight must involve genuine authority. Responsible AI is therefore a continuing organizational capability, not a one-time technical feature.

Key Terms

AI ethics: The study and application of moral principles to AI design, use, and impacts.

Responsible AI: Policies, practices, and controls that align AI with ethical, legal, professional, and societal requirements.

Trustworthy AI: AI that merits justified confidence through reliability, safety, fairness, transparency, security, and accountability.

Socio-technical system: A system whose outcomes arise from interactions among technology, people, organizations, and social context.

Algorithmic accountability: The obligation and capacity to explain, justify, monitor, and remedy algorithmic consequences.

Explainability: The ability to communicate how an AI system or output can be understood by relevant people.

Interpretability: The extent to which a model’s operation can be understood directly.

Fairness: The ethically and contextually appropriate treatment and distribution of benefits, burdens, and errors.

Autonomy: Respect for people’s capacity to make informed and voluntary choices.

Non-maleficence: The obligation to avoid or reduce foreseeable harm.

Beneficence: The obligation to promote genuine benefits.

Dignity: Recognition that people must not be reduced to demeaning or purely instrumental classifications.

Privacy: Appropriate control and protection of personal information and inferences.

Human oversight: Human review that has sufficient competence, information, authority, and independence.

Impact assessment: A structured examination of an AI system’s likely effects, risks, and mitigations.

Model card: Documentation describing a model’s intended uses, performance, limitations, and risks.

Data provenance: Information about the origin, collection, transformation, and use of data.

Red teaming: Deliberate adversarial testing to identify weaknesses, harms, or misuse.

Contestability: The ability of affected people to question, challenge, and seek review of an AI-assisted decision.

AI assurance: The production of evidence that an AI system meets specified requirements throughout its lifecycle.

Review Questions

Short Answer Questions

1. What is meant by responsible AI?

2. Why is AI described as a socio-technical system?

3. Distinguish explainability from interpretability.

4. What are the two essential dimensions of algorithmic accountability?

5. Why is legal compliance not sufficient for ethical AI?

Long Answer Questions

1. Explain the relationship among AI ethics, responsible AI, and trustworthy AI.

2. Describe the major stages of a responsible-AI lifecycle and the controls relevant to each stage.

3. Discuss how autonomy, justice, privacy, dignity, and non-maleficence influence AI design.

4. Distinguish ethical, legal, technical, and managerial responsibilities in an AI project.

5. Explain why meaningful human oversight requires more than placing a human at the end of an automated workflow.

Analytical / Application Questions

1. A bank uses an AI system to reject loan applications. The system is accurate overall but has higher false-negative rates for one demographic group. Analyse the ethical, technical, legal, and managerial questions that should be investigated.

2. An organization argues that publishing its model’s source code provides sufficient transparency. Evaluate this claim using the concepts of explainability, accountability, and contestability.

3. Design a proportionate governance approach for an AI tool that prioritizes patients for specialist appointments. Identify stakeholders, lifecycle risks, documentation, monitoring, and remedies.

22. Multiple Choice Questions with Answers

Question 1

Which statement best captures the socio-technical nature of an AI system?

A. Its outcomes are determined only by the mathematical model.

B. Its outcomes emerge from interactions among data, technology, people, organizations, and context.

C. Its ethical quality depends only on whether the source code is public.

D. Its social effects can be assessed after deployment without examining institutional processes.

Correct Answer: B. Its outcomes emerge from interactions among data, technology, people, organizations, and context.

Question 2

What distinguishes algorithmic accountability from simple transparency?

A. Accountability requires only publishing the training dataset.

B. Accountability requires identifiable actors who can explain, justify, monitor, and remedy consequences.

C. Accountability means that the algorithm itself is legally responsible.

D. Accountability is achieved whenever a model has high accuracy.

Correct Answer: B. Accountability requires identifiable actors who can explain, justify, monitor, and remedy consequences.

Question 3

A model explanation accurately shows that a protected characteristic was replaced by a close proxy. What follows ethically?

A. The use is automatically acceptable because it is explainable.

B. The explanation proves that the model is fair.

C. Explainability has revealed the logic, but the use may still be discriminatory and unacceptable.

D. The model no longer requires human oversight.

Correct Answer: C. Explainability has revealed the logic, but the use may still be discriminatory and unacceptable.

Question 4

Which activity should occur before selecting a machine-learning model?

A. Publishing the model’s final accuracy score.

B. Defining the legitimate purpose, affected groups, and available alternatives.

C. Automating all decisions to maximize efficiency.

D. Removing all domain experts from the process.

Correct Answer: B. Defining the legitimate purpose, affected groups, and available alternatives.

Question 5

Why may a technically accurate benefits-prioritization system still be ethically problematic?

A. Accuracy always conflicts with fairness.

B. The system may encode historical disadvantage and turn a prediction into a consequential administrative decision.

C. Public-sector systems cannot use data.

D. Human review makes every automated system unethical.

Correct Answer: B. The system may encode historical disadvantage and turn a prediction into a consequential administrative decision.

Question 6

Which condition makes human oversight meaningful?

A. The reviewer approves every output automatically.

B. The reviewer has authority, competence, time, and information to challenge the system.

C. The reviewer is informed only after the decision is final.

D. The organization describes the system as human-in-the-loop.

Correct Answer: B. The reviewer has authority, competence, time, and information to challenge the system.

Question 7

Which statement about fairness metrics is most accurate?

A. One metric universally determines whether an AI system is fair.

B. Fairness metrics are irrelevant if aggregate accuracy is high.

C. Different fairness criteria may conflict and must be interpreted in context.

D. Fairness can be guaranteed by excluding protected attributes from the data.

Correct Answer: C. Different fairness criteria may conflict and must be interpreted in context.

Question 8

A manager delays deployment until subgroup testing, impact assessment, and an appeal process are established. Which responsibility is most directly demonstrated?

A. Managerial and organizational responsibility.

B. Only software-maintenance responsibility.

C. Only responsibility for mathematical optimization.

D. No responsibility, because deployment decisions are purely technical.

Correct Answer: A. Managerial and organizational responsibility.

Question 9

What is the primary purpose of an AI impact assessment?

A. To guarantee that no system will ever produce an error.

B. To examine likely effects, affected groups, risks, and possible mitigations.

C. To replace all legal analysis with a numerical score.

D. To prove that an organization has no accountability.

Correct Answer: B. To examine likely effects, affected groups, risks, and possible mitigations.

Question 10

Why should responsible AI include post-deployment monitoring?

A. Because model and social conditions can change, producing drift or new harms.

B. Because testing before deployment is always useless.

C. Because monitoring eliminates the need for documentation.

D. Because users should never be told that a system has limitations.

Correct Answer: A. Because model and social conditions can change, producing drift or new harms.