Research Ethics, Integrity, and Governance

Introduction

Research in artificial intelligence and machine learning can influence employment, healthcare, education, finance, policing, public services, and individual rights. Because research produces knowledge that may be reused in consequential systems, technical competence alone is insufficient. Researchers must also determine whether a study is ethically acceptable, scientifically honest, legally compliant, socially responsible, and properly governed.

Research ethics concerns the principles that guide responsible decisions involving people, communities, animals, data, institutions, and the environment. Research integrity concerns honesty, rigor, transparency, and accountability throughout the research process. Governance provides the structures, policies, review procedures, documentation, and oversight through which ethical and integrity requirements are implemented.

For AI and ML researchers, these responsibilities extend beyond collecting data and publishing results. They include assessing bias in datasets, protecting personal information, explaining model limitations, documenting design decisions, monitoring deployed systems, and ensuring that accountability is not obscured by technical complexity.

Learning Objectives

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

1. Explain the foundations of research ethics, integrity, and governance.

2. Apply informed-consent, privacy, confidentiality, and data-protection principles.

3. Distinguish acceptable authorship from inappropriate credit and contribution practices.

4. Identify plagiarism, fabrication, falsification, and other forms of research misconduct.

5. Recognize and manage conflicts of interest.

6. Evaluate fairness, transparency, accountability, and risks in AI research.

7. Describe the role of institutional review and research-governance processes.

8. Develop a practical ethical plan for an AI or ML research project.

1. Foundations of Ethical Research

1.1 Respect for persons, beneficence, and justice

Three widely used principles provide a foundation for research involving human participants:

• Respect for persons: Individuals should be treated as autonomous decision-makers. Those with limited autonomy require additional protection.
• Beneficence and non-maleficence: Researchers should seek meaningful benefits and minimize foreseeable physical, psychological, economic, social, and informational harms.
• Justice: The benefits and burdens of research should be distributed fairly. Vulnerable groups should not be selected merely because they are accessible or less able to resist participation.

These principles are complemented by scientific validity. A study that cannot answer its research question wastes participants’ time and may expose them to risk without producing useful knowledge. Ethical review therefore considers both participant protection and methodological quality.

1.2 Ethical theories in research decisions

Ethical reasoning may draw on several perspectives:

• Deontological ethics emphasizes duties and rules, such as the duty to obtain consent and tell the truth.
• Consequentialism evaluates likely benefits and harms of an action.
• Virtue ethics focuses on the character of the researcher, including honesty, fairness, courage, and humility.
• Justice-based approaches examine power, inequality, representation, and who benefits from research.
• Care ethics emphasizes relationships, dependency, context, and responsibilities toward vulnerable people and communities.

No single theory resolves every research dilemma. A responsible decision usually combines rules, consequences, professional virtues, and attention to justice.

2. Informed Consent

Informed consent is an ongoing process through which a competent person freely agrees to participate after receiving understandable information about the research. A signature is evidence of the process, not a substitute for it.

Valid consent normally requires:

1. Information: The participant understands the purpose, procedures, duration, risks, benefits, data uses, confidentiality arrangements, compensation, contacts, and withdrawal rights.

2. Comprehension: Information is presented in language and a format appropriate to the participant’s abilities and circumstances.

3. Voluntariness: Participation is free from coercion and undue influence.

4. Capacity: The individual has the legal and cognitive ability to decide, or an authorized representative provides consent where permitted.

AI studies create additional consent questions. Participants should be told if their data will train a model, be combined with other datasets, be shared with external parties, be used for future research, or support commercial development. Consent obtained for one purpose does not automatically authorize unrelated purposes. Researchers should avoid vague statements such as “data may be used for anything,” particularly where participants could not reasonably anticipate the later use.

Consent may be written, electronic, verbal, or recorded, depending on the study and applicable rules. Waiver or alteration of consent may sometimes be approved for minimal-risk secondary research, but researchers must not decide this independently; an authorized ethics body must review the justification.

3. Privacy, Confidentiality, and Data Protection

These concepts are related but distinct.

• Privacy concerns a person’s control over access to themselves, their information, and their activities.
• Confidentiality concerns the researcher’s obligation to protect information entrusted by participants.
• Anonymity means that identity is not collected or cannot reasonably be linked to the data.
• Pseudonymization replaces direct identifiers with codes, while a separate key may permit re-identification.
• Data protection includes the technical, organizational, and legal measures used to govern personal data.

Removing names does not necessarily anonymize an AI dataset. Combinations of age, location, occupation, timestamps, images, writing style, or device information may enable re-identification. Researchers should assess the possibility of linkage and inference rather than relying on a label such as “de-identified.”

A responsible data-management plan should specify:

• what data are collected and why;
• the lawful or institutional basis for processing;
• data minimization and retention periods;
• access permissions and authentication;
• encryption during storage and transmission;
• backup and recovery procedures;
• re-identification risks;
• data-sharing conditions;
• deletion or archival procedures; and
• procedures for reporting breaches.

Security is necessary but not sufficient. A perfectly secure database may still contain data collected without valid consent or used for an unfair purpose. Researchers must consider the entire data life cycle: collection, preparation, storage, analysis, publication, sharing, deployment, and destruction.

4. Research Integrity and Misconduct

Research integrity means conducting and communicating research honestly, accurately, transparently, and responsibly. The major forms of research misconduct are:

• Fabrication: Making up data, participants, experiments, or results.
• Falsification: Manipulating materials, procedures, images, code, or data so that the research record is misrepresented.
• Plagiarism: Presenting another person’s words, ideas, code, images, or results as one’s own without appropriate acknowledgment.

Questionable research practices may not meet the formal definition of misconduct but can still damage scholarship. Examples include selective reporting, undisclosed changes to hypotheses, inappropriate exclusion of observations, p-hacking, duplicate publication, misleading visualizations, inadequate version control, and failure to preserve data or code.

A researcher may legitimately clean data when rules are defined in advance and reported transparently. Removing observations only after seeing which choice produces a preferred result is problematic. Similarly, a model should not be evaluated repeatedly on the test set and then described as having been tested on unseen data.

Good practice includes preregistering important hypotheses where suitable, maintaining a research log, preserving raw data, recording data-processing decisions, using version control, reporting negative results, and retaining enough information for an independent researcher to understand or reproduce the work.

5. Authorship, Contribution, and Publication Ethics

Authorship should reflect substantial intellectual or practical contribution, participation in drafting or critically revising the work, approval of the final version, and willingness to accept responsibility for the contribution. Exact institutional standards may differ, so research teams should agree on criteria at the beginning and revisit them as the project develops.

Common problems include:

• Gift or honorary authorship: Listing someone because of status, supervision, funding, or friendship without substantial contribution.
• Ghost authorship: Omitting a person who made a substantial contribution.
• Coercive authorship: Pressuring junior researchers to add senior individuals.
• Guest authorship: Adding a well-known name to improve credibility.
• Unfair author order: Treating order as a reward without an agreed rationale.

Technical contributions matter. A person who designs a major experiment, develops essential software, creates a dataset, or performs substantial analysis may deserve authorship, subject to the full contribution requirements. Contributors who do not qualify for authorship should be recognized in acknowledgments where appropriate.

Researchers must also avoid duplicate submission, salami publication that fragments one study unnecessarily, undisclosed translation or republication, and manipulation of peer review. Citation should credit relevant prior work rather than merely increase citation counts.

6. Conflict of Interest

A conflict of interest exists when a secondary interest could improperly influence, or reasonably appear to influence, professional judgment. Financial interests include investments, consulting payments, grants, patents, employment, and commercial relationships. Non-financial conflicts may involve personal relationships, academic competition, institutional pressure, political commitments, or strong intellectual commitments.

A conflict is not automatically wrongdoing. The essential requirements are disclosure, assessment, and management. Management may include independent analysis, separation of sponsor and researcher decisions, oversight by an uninvolved person, public disclosure, restriction from a decision, or withdrawal from the project.

For example, a researcher evaluating a model developed by their own start-up should disclose the relationship and consider independent replication. A statement that no conflict exists is inappropriate if a reasonable reader could identify a relevant competing interest.

7. Institutional Review and Research Governance

Institutional review boards, research ethics committees, data-governance offices, and similar bodies assess whether proposed research is ethically and legally acceptable. Review commonly considers:

• the scientific rationale and risk–benefit balance;
• participant recruitment and vulnerability;
• consent materials;
• privacy and confidentiality;
• data security and retention;
• compensation and adverse-event procedures;
• inclusion and exclusion criteria;
• conflicts of interest; and
• plans for dissemination and future data use.

Review may be full, expedited, exempt, or delegated depending on the risk and jurisdiction. “Exempt” does not mean “unregulated,” and researchers should obtain an official determination rather than making an informal assumption.

Approval is not permanent permission to change the study freely. Material changes to participants, procedures, data uses, recruitment, or risk normally require amendment approval. Unexpected harms, breaches, protocol deviations, and serious complaints should be reported through the required institutional channels.

Governance also includes institutional policies, funder conditions, professional codes, contracts, data-sharing agreements, audit trails, training, and mechanisms for whistleblowing and investigation. Good governance should protect participants without creating unnecessary bureaucracy that prevents valuable low-risk research.

8. Responsible AI Research

AI systems can reproduce or amplify social inequalities because of biased data, inappropriate labels, unequal access, historical discrimination, or design choices. Responsible AI is therefore a socio-technical responsibility: fairness cannot be achieved by changing an algorithm alone.

8.1 Fairness

Fairness has multiple meanings. Demographic parity compares selection rates across groups. Equal opportunity compares true-positive rates across groups. Equalized odds compares both true-positive and false-positive rates. Individual fairness asks whether similar individuals receive similar treatment according to a justified similarity measure.

These criteria may conflict, especially when groups have different base rates. No fairness metric should be selected mechanically. Researchers must define the relevant harm, affected population, decision context, and legitimate purpose. Fairness assessment should examine intersectional groups, not only broad categories such as gender or race.

Metrics can hide harms. An apparently accurate model may perform poorly for a small group. Conversely, equal error rates do not necessarily mean that the consequences of errors are equal. Evaluation should combine quantitative performance with stakeholder consultation and domain expertise.

8.2 Transparency and explainability

Transparency includes clear documentation of data sources, labeling procedures, model architecture, evaluation conditions, limitations, intended uses, and prohibited uses. Explainability refers to methods that help people understand model behavior, but an explanation is not automatically faithful merely because it is easy to read.

Useful documentation may include dataset statements, model cards, experiment logs, data sheets, risk assessments, and records of rejected alternatives. Claims should be proportionate to evidence. A model evaluated only on a laboratory dataset should not be described as reliable in clinical or public settings.

8.3 Accountability and human oversight

Accountability requires identifiable responsibility. “The algorithm decided” is not an adequate allocation of responsibility. Researchers and organizations should define who approves deployment, monitors performance, handles complaints, investigates failures, and can suspend or withdraw the system.

Human oversight is meaningful only when the human has sufficient information, authority, time, competence, and independence to challenge the system. A nominal human reviewer who routinely accepts automated outputs without review does not provide effective oversight.

8.4 Generative AI and research

Generative AI tools create risks involving fabricated citations, unverifiable claims, leakage of confidential manuscripts or participant data, copyright concerns, and undisclosed assistance. Researchers should verify every generated claim, avoid entering restricted information into external systems, disclose material use according to institutional and publisher policies, and retain human responsibility for the submitted work. AI tools should not be listed as authors because they cannot accept responsibility, approve a final manuscript, or disclose conflicts.

9. A Practical Ethical Workflow for AI Research

An ethical workflow can be organized into seven stages:

1. Define purpose and necessity: Clarify the research question, intended users, affected groups, and why personal or sensitive data are needed.

2. Map stakeholders and harms: Identify participants, non-participants, data subjects, institutions, downstream users, and potentially excluded groups.

3. Assess risk and governance requirements: Determine whether institutional review, data-protection review, security assessment, or specialist oversight is required.

4. Design safeguards: Minimize data, obtain consent where appropriate, restrict access, establish retention rules, and plan incident response.

5. Document methods and decisions: Record provenance, preprocessing, exclusions, model versions, evaluation protocols, limitations, and conflicts.

6. Evaluate performance and impact: Test accuracy, robustness, subgroup performance, privacy, security, and foreseeable misuse.

7. Monitor and revise: Track real-world outcomes, accept complaints, investigate failures, update safeguards, and stop the system when risks become unacceptable.

This workflow is iterative. Ethical review should not occur only at the moment of application submission or immediately before publication.

10. Illustrative Case Study: A University Early-Warning Model

Background

A university develops an ML model to identify students who may fail a course. The model uses attendance, assessment scores, learning-platform activity, disability-support records, and demographic variables.

Problem

The proposed system could help advisors offer support, but it could also stigmatize students, expose sensitive information, reproduce unequal treatment, or encourage staff to treat predictions as facts. Some students did not expect learning-platform activity to be used for automated risk scoring.

Analysis

The project involves personal data and potentially sensitive information. The research team must clarify whether the model is for research, operational intervention, or both. It must obtain appropriate institutional approval, justify each data field, assess consent and notice, restrict access, and determine whether disability data are necessary. Historical academic outcomes may reflect unequal teaching conditions rather than student ability.

The team should separate training, validation, and final testing; evaluate subgroup error rates; assess calibration; examine false positives and false negatives; and consult students, advisors, accessibility specialists, and legal or governance officers. A prediction should not automatically trigger disciplinary action or deny opportunities.

Approach and solution

The university removes unnecessary sensitive variables, documents the remaining data sources, uses role-based access, and provides students with a clear notice and a channel to ask questions. Advisors receive training that predictions are prompts for supportive conversation, not determinations of ability or entitlement. The system is piloted, independently reviewed, monitored for disparate outcomes, and suspended if harmful patterns appear.

Outcome

The pilot identifies useful indicators but also shows that students working off-campus have lower platform activity despite satisfactory performance. The university therefore revises the model and expands non-technical support routes instead of deploying the original version.

Learning points

1. Ethical risk includes the purpose and consequences of a system, not only its accuracy.

2. Data minimization can improve both privacy and validity.

3. Fairness analysis must consider context and potential causes of observed differences.

4. Human oversight requires authority and training, not merely human presence.

5. A responsible decision may be to revise, limit, or stop deployment.

11. Challenges and Limitations

Ethical frameworks do not eliminate disagreement. Privacy may conflict with reproducibility; fairness may conflict with another fairness criterion; open data may conflict with participant expectations; and transparency may conflict with security or intellectual-property restrictions. Consent can also be difficult when data are collected indirectly or when future uses cannot be fully predicted.

Researchers should acknowledge uncertainty rather than claiming that a system is completely unbiased, private, or explainable. Risk assessment can be limited by unknown downstream uses, changing models, incomplete demographic information, and power differences between institutions and participants. Governance bodies may lack technical expertise, while technical teams may underestimate social and legal consequences. These limitations make interdisciplinary review and continuing monitoring essential.

12. Best Practices for Students and Research Teams

• Complete ethics and data-protection training before collecting data.
• Ask for institutional guidance before assuming that a project is exempt.
• Use the minimum data necessary for the stated purpose.
• Obtain meaningful consent and explain secondary uses.
• Separate identifiers from research data and protect the linking key.
• Maintain reproducible records of code, data transformations, and decisions.
• Agree on authorship and contribution expectations early.
• Declare financial and non-financial conflicts.
• Cite words, ideas, code, datasets, and images accurately.
• Report limitations, null findings, and unexpected results.
• Evaluate AI systems across relevant and intersectional groups.
• Define responsibility, escalation routes, and stop conditions.
• Protect confidential information from unapproved AI tools.
• Treat ethics as a continuing process rather than a form completed once.

Chapter Summary

Research ethics protects people, communities, and the integrity of knowledge. Respect for persons, beneficence, non-maleficence, and justice guide decisions about participation and risk. Informed consent must be understandable, voluntary, and appropriate to the proposed data uses. Privacy, confidentiality, anonymity, pseudonymization, and data protection are distinct concepts requiring life-cycle safeguards.

Research integrity requires honest methods, accurate records, appropriate credit, responsible publication, and rejection of fabrication, falsification, and plagiarism. Authorship should reflect substantial contribution and responsibility. Conflicts of interest should be disclosed and managed rather than concealed. Institutional review and governance provide structured oversight, but approval does not remove the researcher’s continuing responsibility.

In AI and ML, responsible research requires attention to fairness, transparency, explainability, accountability, privacy, security, and human oversight. Metrics alone cannot settle ethical questions. A practical workflow connects purpose, stakeholder analysis, risk assessment, safeguards, documentation, evaluation, and monitoring. Ethical research is therefore not an obstacle added to technical work; it is part of producing valid, trustworthy, and socially useful knowledge.

Key Terms

Research ethics: Principles governing responsible research involving people, data, institutions, and society.

Research integrity: Honesty, rigor, transparency, and accountability in conducting and reporting research.

Informed consent: Voluntary agreement based on adequate information and understanding.

Privacy: Control over access to a person, their activities, or their information.

Confidentiality: Duty to protect information entrusted to the researcher.

Anonymity: Condition in which identity cannot reasonably be linked to research data.

Pseudonymization: Replacement of identifiers with codes while preserving possible re-identification through a separate key.

Data minimization: Collection and use of no more personal data than necessary.

Fabrication: Creation of false data, participants, procedures, or results.

Falsification: Manipulation that misrepresents research materials, methods, or findings.

Plagiarism: Use of another’s work or ideas without appropriate acknowledgment.

Authorship: Formal credit and responsibility assigned to qualifying contributors.

Conflict of interest: A secondary interest that could influence or appear to influence judgment.

Institutional review: Formal assessment of a research project by an authorized ethics or governance body.

Fairness: Justified and equitable treatment of affected individuals or groups.

Transparency: Clear disclosure of relevant data, methods, assumptions, limitations, and decisions.

Accountability: Identification of people or organizations responsible for decisions and outcomes.

Human oversight: Meaningful human ability and authority to evaluate, challenge, and intervene in automated decisions.

Responsible AI: Development and use of AI consistent with ethical, legal, social, and professional responsibilities.

Review Questions

Short Answer Questions

1. Distinguish privacy, confidentiality, anonymity, and pseudonymization.

2. What are the essential elements of informed consent?

3. Define fabrication, falsification, and plagiarism.

4. Why should conflicts of interest be disclosed even when no misconduct has occurred?

5. What is the purpose of institutional research review?

Long Answer Questions

1. Explain how respect for persons, beneficence, and justice apply to an AI research project.

2. Discuss the ethical and practical requirements of a data-management plan.

3. Explain responsible authorship and describe common authorship abuses.

4. Compare transparency, explainability, and accountability in AI research.

5. Describe how research integrity can be supported through documentation and reproducible practice.

Analytical / Application Questions

1. A researcher wants to combine public social-media posts with university records to predict student dropout. Identify the major ethical, privacy, consent, governance, and fairness issues and propose safeguards.

2. An image-classification model has high overall accuracy but substantially higher false-negative rates for one demographic group. Analyze possible causes and recommend an evaluation and remediation plan.

3. A senior supervisor demands authorship despite making no intellectual contribution. Apply authorship and integrity principles to develop a professional response and escalation strategy.

Multiple Choice Questions with Answers

Question 1

Which principle requires that the burdens and benefits of research be distributed fairly?

A. Respect for persons

B. Justice

C. Confidentiality

D. Transparency

Correct Answer: B. Justice

Question 2

Which statement best describes pseudonymization?

A. Data are collected without any identifiers

B. Data are permanently destroyed after analysis

C. Identifiers are replaced by codes but re-identification may remain possible

D. Data are made publicly accessible

Correct Answer: C. Identifiers are replaced by codes but re-identification may remain possible

Question 3

A researcher invents participant responses that were never collected. This is:

A. Fabrication

B. Falsification

C. Plagiarism

D. Pseudonymization

Correct Answer: A. Fabrication

Question 4

Which practice most directly supports research integrity?

A. Deleting inconvenient results without explanation

B. Changing the hypothesis after seeing results and concealing the change

C. Maintaining records of data processing and model versions

D. Reporting only the best-performing subgroup

Correct Answer: C. Maintaining records of data processing and model versions

Question 5

What is the primary purpose of informed consent?

A. To guarantee that the research will benefit every participant

B. To obtain voluntary agreement based on understandable information

C. To transfer all participant rights to the researcher

D. To eliminate the need for institutional review

Correct Answer: B. To obtain voluntary agreement based on understandable information

Question 6

Which is an appropriate response to a conflict of interest?

A. Conceal it if the results appear accurate

B. Disclose and manage it using appropriate safeguards

C. Automatically cancel every study involving a conflict

D. Transfer authorship to the sponsor

Correct Answer: B. Disclose and manage it using appropriate safeguards

Question 7

Why may removing names fail to anonymize an AI dataset?

A. Names are the only possible personal data

B. Other attributes may enable linkage or re-identification

C. Anonymization requires publishing all records

D. Machine-learning models cannot use anonymous data

Correct Answer: B. Other attributes may enable linkage or re-identification

Question 8

Which statement about human oversight is most accurate?

A. Any human presence automatically makes a system accountable

B. Human oversight requires authority, competence, information, and ability to intervene

C. Human oversight is unnecessary for high-impact systems

D. Human oversight means accepting every model recommendation

Correct Answer: B. Human oversight requires authority, competence, information, and ability to intervene

Question 9

A model has equal accuracy across groups but different error consequences. What should researchers do?

A. Conclude immediately that the model is fair

B. Ignore consequences because accuracy is sufficient

C. Examine relevant harms, error types, context, and stakeholder impacts

D. Remove all demographic information without further analysis

Correct Answer: C. Examine relevant harms, error types, context, and stakeholder impacts

Question 10

Why should AI tools not be treated as authors of a research paper?

A. They cannot accept responsibility, approve the final work, or disclose conflicts

B. They cannot produce text in academic language

C. They are always less accurate than humans

D. They cannot be used in research under any circumstances

Correct Answer: A. They cannot accept responsibility, approve the final work, or disclose conflicts