Chapter 14: Ethics, Reproducibility and Responsible Research

1. Introduction

Quantitative and qualitative data analysis does not occur in an ethical vacuum. Every dataset represents people, institutions, communities, environments, or decisions that may be affected by research. The collection, storage, interpretation, and publication of data therefore involve questions of power, privacy, accountability, and justice. At doctoral level, methodological competence must be accompanied by the ability to identify ethical risks, justify analytical decisions, document research processes, and communicate findings responsibly.

Ethical research is not limited to obtaining approval from an institutional review board or research ethics committee. It includes designing proportionate data practices, obtaining meaningful consent, protecting confidentiality, recognizing community rights, managing conflicts of interest, testing algorithms for bias, and ensuring that conclusions are not stronger than the evidence. Reproducibility adds a related obligation: researchers should make it possible for qualified others to understand, inspect, and, where appropriate, reproduce the path from data and code to findings.

This chapter presents an integrated framework for ethics, reproducibility, and responsible research in information technology. It addresses conventional research involving human participants as well as contemporary problems involving surveillance, artificial intelligence, platform data, automated decision systems, and Indigenous or community-held knowledge.

2. Learning Objectives

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

1. Explain the ethical foundations of quantitative and qualitative data analysis.

2. Distinguish informed consent, privacy, confidentiality, anonymity, and data sovereignty.

3. Identify risks arising from surveillance, algorithmic bias, conflicts of interest, and secondary data use.

4. Design audit trails, version-control practices, and reproducible analytical workflows.

5. Apply fairness, transparency, accountability, and human-oversight principles to responsible AI.

6. Evaluate open-science practices in relation to privacy, intellectual property, and community rights.

7. Develop governance strategies for ethically responsible doctoral research.

3. Meaning and Concept

Research ethics is the systematic application of moral principles to the design, conduct, analysis, and communication of research. It asks whether a study respects persons, minimizes harm, distributes benefits and burdens fairly, and remains honest and accountable.

Reproducibility refers to the capacity of another researcher to obtain the same or substantially similar results using the documented materials, data, code, and procedures. In a strict sense, reproducibility often means rerunning the same analysis with the same data. Replicability usually refers to obtaining consistent findings in a new study using new data and an equivalent design. Both concepts are important, but neither means that every study must produce identical results. Qualitative research may instead emphasize transparency, interpretive adequacy, reflexivity, and confirmability.

Responsible research integrates ethical judgment, methodological rigor, social awareness, and institutional accountability. It recognizes that technically possible data use is not automatically morally permissible. For example, a researcher may be able to scrape public social-media posts, but public accessibility does not necessarily eliminate expectations of privacy or the possibility of harm through quotation, profiling, or re-identification.

4. Fundamental Principles

The Belmont principles remain influential in human-subject research. Respect for persons requires recognition of autonomy and additional protection for people with limited capacity to consent. Beneficence requires maximizing anticipated benefits and minimizing possible harm. Justice requires fair selection of participants and fair distribution of research burdens and benefits.

These principles can be extended through several operational commitments. Integrity means reporting methods and results honestly, including null findings, exclusions, failed models, and uncertainty. Transparency means making decisions understandable to participants, reviewers, users, and affected communities. Accountability means assigning responsibility for data and model outcomes rather than treating an algorithm as an independent actor. Proportionality requires that data collection and intrusion be no greater than necessary for the research purpose.

Fairness is both a technical and social principle. A model may satisfy one statistical fairness criterion while producing unjust outcomes because the target variable, sampling frame, or institutional policy is itself discriminatory. Fairness analysis therefore requires attention to context, history, and consequences, not only mathematical parity.

5. Major Components / Elements

Informed consent and autonomy

Informed consent is a process, not merely a signed form. Participants should understand the purpose of the study, procedures, foreseeable risks, expected benefits, data uses, withdrawal conditions, compensation, contacts for concerns, and plans for retention or destruction. Consent materials should be comprehensible and translated or adapted where necessary. Researchers must avoid undue influence, especially where participation is linked to employment, grades, healthcare, immigration status, or access to services.

Broad or secondary-use consent may be appropriate in some longitudinal or repository studies, but it cannot be treated as unlimited permission. New uses should be compatible with the original agreement, legally permissible, and ethically reviewed. When re-consent is impractical, researchers should document why, assess risk, and apply additional safeguards.

Privacy, confidentiality, and anonymity

Privacy concerns a person’s control over access to themselves and their information. Confidentiality concerns the researcher’s duty to protect information entrusted to them. Anonymity means that identity cannot reasonably be linked to the data; it is often difficult to guarantee in longitudinal, geospatial, or small-community research. Removing names is insufficient when combinations of age, location, occupation, dates, and rare characteristics permit re-identification.

Safeguards include data minimization, pseudonymization, aggregation, encryption, access controls, secure transfer, retention limits, and disclosure review. Researchers should assess both direct and inferential identification risks. A dataset can be formally de-identified yet allow sensitive attributes to be inferred from correlated variables or external databases.

Surveillance and the datafication of life

Digital cameras, biometric systems, workplace monitoring, location services, learning analytics, and online tracking can generate data without meaningful awareness or choice. Surveillance research should examine necessity, proportionality, function creep, power asymmetries, and the possibility that observation changes behavior. Institutional permission to access a system does not automatically authorize intrusive analysis of individuals.

A responsible design identifies the least intrusive means of answering the research question. It also specifies who can access outputs, whether individual-level intervention is possible, how errors will be corrected, and whether people can contest decisions based on collected data.

Algorithmic bias and fairness

Bias may enter through problem formulation, sampling, measurement, labels, missing data, historical practices, model architecture, deployment, or feedback loops. For example, using prior arrest records as a measure of criminality reproduces patterns of policing rather than directly measuring offending. A model trained on one population may perform poorly for another because of distributional differences.

Fairness assessment should include subgroup performance, error types, calibration, intersectional analysis, and impact after deployment. No single fairness metric is universally correct. Researchers must justify the chosen criterion and identify trade-offs. Qualitative inquiry with affected stakeholders is often necessary to reveal harms that aggregate performance statistics conceal.

Indigenous and community data rights

Indigenous peoples and local communities may possess collective rights over data, cultural knowledge, language resources, biological materials, and narratives. Individual consent alone may be inadequate where information has collective meaning or consequences. Principles such as Indigenous data sovereignty and the CARE principles—Collective benefit, Authority to control, Responsibility, and Ethics—encourage governance that recognizes community authority and benefit.

Researchers should engage communities early, define ownership and access arrangements, respect local protocols, return useful findings, and avoid extracting knowledge without reciprocal benefit. Some data should not be openly published even when dissemination would improve scientific discoverability. Community governance may require review, licensing, data access committees, or restrictions on secondary use.

Conflicts of interest and research integrity

A conflict of interest exists when personal, financial, professional, political, or institutional interests could reasonably appear to influence research judgment. Conflicts are not necessarily misconduct; undisclosed or poorly managed conflicts are the central concern. Researchers should disclose funding, consultancy, equity, supervisory relationships, incentives, and relevant affiliations.

Integrity also requires avoiding fabrication, falsification, plagiarism, deceptive image manipulation, inappropriate authorship, duplicate publication, p-hacking, HARKing, and selective reporting. Pre-registration, analysis plans, independent review, and full documentation reduce opportunities for undisclosed flexibility.

6. Process / Framework / Methodology

A practical responsible-research framework can be organized into eight stages.

First, define the purpose and affected parties. State the research question, intended users, foreseeable beneficiaries, and groups that could be harmed. Ask whether the question can be answered with less sensitive data.

Second, map the data lifecycle. Document collection, transfer, transformation, storage, analysis, sharing, archiving, and destruction. Identify controllers, processors, custodians, and authorized users.

Third, conduct an ethical and impact assessment. Evaluate consent, privacy, security, discrimination, re-identification, environmental costs, community rights, and possible misuse. Include stakeholder consultation rather than relying solely on an internal checklist.

Fourth, design governance controls. Establish access permissions, retention periods, encryption, incident-response procedures, withdrawal processes, model documentation, and escalation routes. Assign named responsibility.

Fifth, create a reproducible workflow. Use a structured repository, version control, scripted transformations, documented environments, data dictionaries, and a research log. Separate raw data from processed data and preserve immutable copies where lawful and safe.

Sixth, analyze with reflexivity and robustness. Quantitative researchers should record preprocessing, missing-data decisions, exclusions, model specifications, sensitivity analyses, and uncertainty. Qualitative researchers should document sampling, coding decisions, researcher positionality, memoing, negative cases, and changes to the interview guide.

Seventh, validate and audit. Use peer review, code review, independent reanalysis, subgroup testing, member reflection where appropriate, privacy checks, and adversarial testing. Validation should include both technical performance and social consequences.

Eighth, communicate and monitor. Report limitations, uncertainties, conflicts, and residual risks. Establish post-publication correction, incident reporting, model monitoring, and data-access review. Responsibility continues after publication or deployment.

7. Important Theories and Models

The principlism model uses respect for autonomy, beneficence, non-maleficence, and justice to evaluate competing obligations. It is flexible but may require additional reasoning when principles conflict.

The privacy-by-design model embeds privacy into system architecture rather than adding it after development. Typical principles include proactive prevention, default privacy, full functionality, end-to-end security, visibility, and respect for user interests.

The FAIR principles recommend that research outputs be Findable, Accessible, Interoperable, and Reusable. FAIR does not mean unrestricted public release. Access may be controlled to protect privacy, safety, intellectual property, or community authority.

The CARE principles complement FAIR by emphasizing collective rights and community benefit. They are particularly relevant to Indigenous and community data governance.

The Responsible Research and Innovation model encourages anticipation of impacts, reflection on assumptions, inclusion of affected groups, and responsiveness to emerging evidence. It treats innovation as a social process rather than a purely technical achievement.

A sociotechnical systems perspective holds that outcomes arise from interactions among people, technologies, organizational procedures, incentives, and law. It prevents researchers from attributing responsibility solely to data or algorithms.

8. Practical Applications

In a quantitative study, a researcher might preregister hypotheses, define exclusion criteria before examining outcomes, use a reproducible script, report missingness, and conduct sensitivity analyses. If a dataset includes rare diseases or small geographic areas, access may be provided through a secure enclave rather than a public repository.

In qualitative research, an audit trail can include recruitment decisions, consent adaptations, interview revisions, coding frameworks, reflexive memos, disagreements among analysts, and reasons for merging or separating themes. Direct quotations should be checked for deductive disclosure, particularly in small organizations or communities.

For machine-learning research, a model card can describe intended use, training data, limitations, evaluation groups, known risks, and prohibited uses. A data sheet can document dataset motivation, composition, collection process, preprocessing, recommended uses, and ethical concerns. These artifacts support informed scrutiny by users and affected parties.

9. Examples

A university analyzes learning-management-system logs to predict student withdrawal. Ethical questions include whether students were informed, whether the model will support or penalize them, whether disability or socioeconomic status is indirectly inferred, and whether advisors can challenge an automated risk score. A responsible system uses minimum necessary data, provides supportive rather than punitive intervention, evaluates subgroup error rates, and communicates that a prediction is uncertain.

A researcher studies public posts about domestic violence. Although the posts are visible online, quoting them verbatim may expose authors through search engines. The researcher should consider paraphrasing, removing metadata, seeking consent where feasible, and presenting aggregate themes rather than searchable quotations.

A community health project collects genetic and environmental data from a small Indigenous population. Individual consent, community approval, data sovereignty agreements, benefit sharing, culturally appropriate interpretation, and restrictions on commercial secondary use are all necessary. Open publication may be limited by community governance.

10. Technology and Digital Applications

Version-control systems support traceable changes to code, documentation, and analysis plans. Repositories should use meaningful commit messages, branch protection, code review, and tagged releases. Sensitive data should not be placed in public repositories; secrets and credentials must be excluded.

Workflow tools can automate data validation, transformation, statistical analysis, and report generation. Containers or environment-locking files help preserve software dependencies. Continuous integration can run tests whenever code changes, reducing accidental analytical errors.

Electronic laboratory notebooks, secure institutional storage, encrypted backups, role-based access, and audit logs support chain-of-custody and accountability. Metadata standards and persistent identifiers improve discovery and citation. Privacy-enhancing technologies, including differential privacy, secure multiparty computation, federated learning, and synthetic data, can reduce exposure, although each introduces assumptions and possible utility loss.

Responsible AI systems may include explainability interfaces, fairness dashboards, model cards, data documentation, human review, drift monitoring, and appeal mechanisms. Explainability should not be confused with justification: a technically interpretable model can still be based on an unjust objective.

11. Strategic / Managerial Importance

Ethical and reproducible research protects participants, institutions, research funding, and public trust. It reduces legal, reputational, operational, and scientific risks. For managers, governance creates clarity about who may collect data, who approves secondary use, who responds to incidents, and who is accountable for automated decisions.

Reproducibility improves efficiency because documented workflows reduce duplicated effort and make handovers possible. It also strengthens strategic decision-making by distinguishing robust evidence from fragile or selectively reported results. Organizations that integrate ethics into procurement, data architecture, staff training, performance evaluation, and project review are better positioned to innovate responsibly.

12. Detailed Case Study

Background

The following is a fictional illustrative case. Northbridge University develops an AI system to identify undergraduate students who may require academic support. The system uses attendance records, learning-platform activity, assessment submissions, and historical advising data.

Problem

The university intends to contact high-risk students early. However, students were not clearly informed that their activity data would be used for predictive profiling. Initial testing shows higher false-positive rates for international students and students who work off campus. Advisors begin treating risk scores as facts, and one student loses access to a scholarship after an administrator informally interprets a score as evidence of poor commitment.

Analysis

The problem is not simply model accuracy. The research team identifies weak consent, excessive data collection, proxy discrimination, inadequate documentation, function creep, and absent appeal procedures. Historical advising data also reflect unequal access to support and therefore provide a biased outcome label. The team discovers that code changes were not versioned and that an undocumented preprocessing step excluded students with incomplete records.

Strategy / Solution

The university pauses high-stakes use and establishes a governance group including students, advisors, disability specialists, data-protection officers, and independent researchers. It revises the purpose from punitive prediction to voluntary support. Data are minimized, sensitive proxies are assessed, and subgroup performance is evaluated across intersecting characteristics where sample sizes permit.

The team creates a data dictionary, model card, preregistered evaluation plan, version-controlled repository, reproducible pipeline, and audit trail. A human advisor must review evidence before contact, students receive an explanation and may correct records or decline participation, and no score can determine funding or disciplinary action. The system is monitored for drift and adverse outcomes. Results are reported with uncertainty and limitations rather than a single accuracy figure.

Outcome

The revised program produces fewer unnecessary interventions and improves student trust. Some predictive performance decreases because unnecessary variables are removed, but the university considers the reduction acceptable. Independent review finds that support is more evenly offered across student groups. The university also learns that some students preferred direct access to advising rather than algorithmic screening.

Learning Points

The case demonstrates that responsible AI is a lifecycle responsibility. Consent, purpose limitation, fairness, human judgment, documentation, appeal, and monitoring must be designed together. A technically accurate model can still be ethically unacceptable if its use undermines autonomy or distributes harm unfairly.

13. Advantages / Benefits

Ethically governed and reproducible research improves credibility, auditability, validity, participant protection, and institutional learning. It enables independent verification, facilitates collaboration, supports correction of errors, and makes research outputs more reusable. Fairness assessment can reveal weaknesses in data collection and service delivery. Community participation can improve relevance, interpretation, and benefit distribution. Transparent limitations also produce more realistic managerial expectations.

14. Limitations / Challenges

Reproducibility requires time, skills, infrastructure, and resources. Some data cannot be openly shared because of privacy, contractual, legal, security, or community obligations. Repositories may become obsolete, software dependencies may disappear, and anonymization may be imperfect. Fairness metrics can conflict, and small subgroup samples may make estimates unstable. Consent may not anticipate future technologies. Ethics committees may lack specialized expertise in AI, Indigenous rights, cybersecurity, or qualitative methods. Finally, transparency itself can create risk if detailed disclosure enables re-identification or system gaming.

15. Ethical and Governance Considerations

Governance should distinguish accountability from ownership. A data custodian may manage access, while researchers, institutions, vendors, and decision-makers share responsibility for consequences. Contracts should specify permitted uses, security requirements, audit rights, breach notification, deletion, subcontracting, and rights of affected individuals.

Organizations should establish independent review for high-risk research, maintain conflict-of-interest registers, protect whistleblowers, and provide accessible complaint and appeal channels. Data protection should be risk-based and continuous. Researchers should not assume that ethics approval transfers all responsibility; professional judgment remains necessary when circumstances change.

Publication decisions should consider whether releasing code, data, prompts, or model weights creates foreseeable harm. When full openness is inappropriate, researchers can provide controlled access, synthetic examples, aggregate statistics, executable containers without sensitive data, or detailed methodological documentation. Ethical restrictions should be explained rather than used as a vague excuse for opacity.

16. Current and Emerging Trends

Open science is moving toward preregistration, registered reports, open methods, open code, transparent peer review, and data stewardship. At the same time, privacy-preserving computation is enabling analysis across institutions without centralizing raw data. Generative AI is creating new issues involving fabricated content, undisclosed assistance, training-data provenance, copyright, confidential information, and reproducibility of stochastic outputs.

Regulatory attention is increasing around automated decision systems, high-risk AI, biometric identification, and data protection. Algorithmic impact assessments, public-sector transparency registers, and independent audits are becoming more common. Digital twins, multimodal datasets, neurotechnology, and synthetic data will require renewed attention to consent, identity, and inference. Community-led governance and data trusts may provide alternatives to purely institutional control.

17. Best Practices

1. Define a legitimate, specific, and proportionate research purpose.

2. Involve affected participants and communities before finalizing the design.

3. Obtain meaningful consent and revisit it when uses materially change.

4. Minimize data collection and protect sensitive information throughout its lifecycle.

5. Separate identifiers from analytical data and control access by role.

6. Use preregistration or an analysis plan when appropriate, while documenting justified deviations.

7. Keep raw, processed, and derived data clearly separated.

8. Use version control, automated tests, environment documentation, and reproducible scripts.

9. Maintain an audit trail for analytical and interpretive decisions.

10. Report missing data, exclusions, uncertainty, null findings, and conflicts of interest.

11. Test subgroup and intersectional performance before deploying models.

12. Provide human oversight, explanations, correction, and appeal mechanisms.

13. Respect Indigenous and community authority over data and knowledge.

14. Share materials according to the principle “as open as possible, as closed as necessary.”

15. Monitor outcomes after publication or deployment and correct errors promptly.

18. Future Scope

Future research will need to integrate technical reproducibility with social reproducibility: the ability to understand not only how a result was computed but also whose assumptions, interests, and institutional conditions shaped it. Automated provenance systems may record data transformations, model versions, and decision pathways, but automation will not replace ethical judgment.

Doctoral researchers are likely to work with dynamic, multimodal, and continuously collected data. This will require adaptive consent, continuous governance, privacy-preserving analytics, and new methods for auditing foundation models. Greater attention will also be given to environmental sustainability, labor conditions in data production, community benefit, and the rights of people affected by inferred rather than explicitly collected data.

19. Chapter Summary

Ethics, reproducibility, and responsible research are mutually reinforcing dimensions of high-quality data analysis. Ethical practice protects autonomy, privacy, welfare, justice, and community rights. Reproducible practice documents the pathway from evidence to conclusion and supports scrutiny and correction. Responsible AI requires attention to purpose, data provenance, bias, fairness, human oversight, accountability, and post-deployment monitoring. Open science should be pursued carefully, recognizing that openness must be balanced with privacy, security, intellectual property, and collective data rights. A doctoral researcher is responsible not only for producing a technically competent analysis but also for anticipating consequences and maintaining an auditable, honest, and socially legitimate research process.

20. Key Terms

1. Audit trail: A chronological record of research decisions, changes, evidence, and approvals.

2. Algorithmic bias: Systematic and unfair differences in automated outputs or impacts.

3. Anonymity: A condition in which identity cannot reasonably be linked to data.

4. CARE principles: Collective Benefit, Authority to Control, Responsibility, and Ethics for community data governance.

5. Confidentiality: The obligation to protect information provided in trust.

6. Conflict of interest: An interest that could influence, or appear to influence, research judgment.

7. Data minimization: Collecting and retaining only data necessary for a legitimate purpose.

8. Data sovereignty: The authority of people or communities to govern data about them.

9. Differential privacy: A mathematical method for limiting disclosure about individuals in statistical outputs.

10. FAIR principles: Findable, Accessible, Interoperable, and Reusable research-data principles.

11. Informed consent: Voluntary agreement based on adequate understanding of research and its risks.

12. Model card: Documentation describing a model’s purpose, performance, limitations, and risks.

13. Open science: Practices that make research methods, outputs, and processes more transparent and accessible.

14. Privacy by design: Integrating privacy protections into system design from the beginning.

15. Reproducibility: The ability to obtain results using the same documented data, code, and procedures.

16. Responsible AI: AI developed and used with attention to fairness, safety, accountability, transparency, and human rights.

17. Surveillance: Systematic observation or monitoring of people, activities, or environments.

18. Version control: A system for recording and managing changes to files, code, and research materials.

21. Review Questions

Short Answer Questions

1. Distinguish privacy, confidentiality, and anonymity.

2. Why is informed consent considered a process rather than a single signature?

3. Explain how version control supports reproducible research.

4. What are two ways that algorithmic bias can enter a research project?

5. Why might data be FAIR without being publicly accessible?

Long Answer Questions

1. Discuss how the Belmont principles apply to contemporary AI-based research.

2. Explain the components of a reproducible quantitative or qualitative workflow.

3. Evaluate the ethical challenges of using public social-media data in research.

4. Discuss the relationship between Indigenous data sovereignty and open science.

5. Design a governance framework for a high-risk automated decision system.

Analytical / Application Questions

1. A hospital wants to predict readmission using electronic health records. Propose a consent, privacy, fairness, audit, and monitoring plan.

2. A qualitative researcher discovers that a quotation in a published draft could identify a participant in a small community. Analyze the options and recommend a course of action.

3. An algorithm has equal overall accuracy across two groups but a much higher false-positive rate for one group. Explain why this matters and propose further investigation and corrective actions.

22. Multiple Choice Questions with Answers

Question 1

Which practice most directly supports reproducibility?

A. Deleting intermediate files after analysis

B. Using version-controlled code and a documented computational environment

C. Reporting only statistically significant findings

D. Replacing missing values without recording the method

Correct Answer: B. Using version-controlled code and a documented computational environment

Question 2

Why is removing names from a dataset insufficient to guarantee anonymity?

A. Names are never useful identifiers

B. Researchers cannot store unnamed data

C. Combinations of variables may permit re-identification

D. Anonymity applies only to qualitative research

Correct Answer: C. Combinations of variables may permit re-identification

Question 3

Which action best reflects meaningful informed consent?

A. Obtaining a signature without explaining secondary data use

B. Allowing participation only when services depend on it

C. Explaining procedures, risks, data uses, and withdrawal conditions clearly

D. Providing information only after data collection

Correct Answer: C. Explaining procedures, risks, data uses, and withdrawal conditions clearly

Question 4

What is a central concern when using historical arrest records as a machine-learning label?

A. The data are always too small to analyze

B. The label may reproduce biased policing patterns rather than actual offending

C. Historical data cannot be stored digitally

D. Labels are unrelated to model fairness

Correct Answer: B. The label may reproduce biased policing patterns rather than actual offending

Question 5

What do the CARE principles primarily emphasize?

A. Faster software deployment

B. Collective benefit and community authority over data

C. Universal public release of all datasets

D. Elimination of qualitative methods

Correct Answer: B. Collective benefit and community authority over data

Question 6

Which response is most appropriate when sensitive data cannot be openly shared?

A. Publish the raw data regardless of risk

B. Abandon all documentation

C. Use controlled access or privacy-preserving alternatives and explain the restriction

D. Share data only with unnamed researchers by email

Correct Answer: C. Use controlled access or privacy-preserving alternatives and explain the restriction

Question 7

Why should algorithmic systems include appeal mechanisms?

A. To guarantee that every prediction is correct

B. To allow affected people to challenge errors and decisions

C. To eliminate the need for human oversight

D. To conceal model limitations

Correct Answer: B. To allow affected people to challenge errors and decisions

Question 8

Which item belongs in a qualitative research audit trail?

A. Only the final list of themes

B. Researchers’ coding decisions, reflexive memos, and reasons for revisions

C. Unverified participant identities

D. A list of unrelated publications

Correct Answer: B. Researchers’ coding decisions, reflexive memos, and reasons for revisions

Question 9

What is the principle of data minimization?

A. Collecting every potentially useful variable

B. Collecting and retaining only data necessary for a legitimate purpose

C. Deleting all data immediately after collection

D. Sharing data with the maximum number of users

Correct Answer: B. Collecting and retaining only data necessary for a legitimate purpose

Question 10

A model has equal overall accuracy across groups but a higher false-positive rate for one group. What is the best conclusion?

A. The model is automatically fair because accuracy is equal

B. The difference is irrelevant if the model is automated

C. Further fairness analysis is needed because error burdens are unequal

D. The disadvantaged group should be excluded from evaluation

Correct Answer: C. Further fairness analysis is needed because error burdens are unequal