4. Data Ethics, Privacy and Intellectual Property

Introduction and Context

Artificial intelligence systems are often described as models, algorithms, or computational infrastructures. In practice, however, their behaviour is strongly shaped by data: what is collected, from whom, under what conditions, for which purpose, with what permissions, and through whose labour. Data is not a neutral input. It is produced through social relations, institutional decisions, technical systems, and legal arrangements. Consequently, ethical AI requires more than accurate models. It requires responsible treatment of people, communities, creators, organisations, and the knowledge embodied in data.

Data ethics concerns the moral principles governing the collection, use, sharing, analysis, retention, and deletion of data. Privacy concerns an individual’s or group’s ability to control information about themselves and to participate in decisions about how that information is used. Intellectual property concerns legally recognised interests in creations, inventions, databases, software, brands, and other forms of knowledge. These domains overlap but are not identical. A dataset may be legally obtainable yet ethically inappropriate to use. A person may have consented to one purpose but not to a later purpose. A work may be publicly accessible but still protected by copyright.

For AI professionals, these issues arise throughout the lifecycle: problem formulation, data acquisition, labelling, model training, evaluation, deployment, monitoring, and retirement. Poor decisions at an early stage can become difficult to correct later. A governance programme must therefore treat provenance, consent, representativeness, privacy, surveillance, ownership, and accountability as design requirements rather than as paperwork added after development.

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the relationship among data ethics, privacy, intellectual property, and AI governance.

2. Evaluate consent, data provenance, ownership, representativeness, and purpose limitation.

3. Distinguish anonymization, pseudonymization, aggregation, and other privacy approaches.

4. Select suitable privacy-enhancing technologies for different AI use cases.

5. Analyse surveillance and data-minimization risks in organisational settings.

6. Identify copyright and licensing issues in datasets, software, models, and generated outputs.

7. Design lifecycle controls for responsible data governance.

8. Apply these ideas to professional and managerial decisions.

Data Is Not Merely a Technical Resource

A useful starting point is to ask four questions about any dataset: whose data is it, how was it produced, what does it represent, and who may benefit or be harmed by its use? These questions expose assumptions hidden by the phrase “the data.” Personal data may describe identifiable individuals, but data can also affect groups, communities, workers, or cultural traditions without identifying particular people. Location traces, purchasing records, biometric signals, health information, educational records, photographs, text, and behavioural logs all carry different levels of sensitivity and different expectations of privacy.

Data ethics extends beyond legal compliance. Law establishes enforceable duties and rights, while ethics also asks whether an action is fair, respectful, proportionate, and socially legitimate. Compliance is therefore a floor, not a complete ethical standard. A firm may comply with a notice-and-consent requirement while presenting the notice in language that ordinary users cannot understand. It may obtain consent for broad “research” while using the data for employment surveillance. It may technically anonymize a dataset while releasing enough contextual information for re-identification.

A useful ethical analysis considers:

• Autonomy: Can people make an informed and voluntary decision about data use?
• Beneficence and non-maleficence: Does the use produce worthwhile benefits while reducing foreseeable harm?
• Justice: Are benefits and burdens distributed fairly, including across marginalised groups?
• Accountability: Can decisions be traced to responsible people and institutions?
• Respect and dignity: Does the system treat people as participants with rights rather than merely as data sources?

These principles can conflict. More data may improve predictive performance but increase privacy risk. A highly customised service may require extensive behavioural monitoring. A dataset balanced for statistical fairness may require collecting sensitive attributes that an organisation would otherwise avoid storing. Governance is therefore the disciplined management of justified trade-offs, not the pursuit of a single universal rule.

Consent, Context, and Legitimate Use

Consent is meaningful only when it is informed, specific enough to be understood, freely given, and capable of being withdrawn where appropriate. It is not equivalent to clicking an “I agree” button. Effective consent communication should identify the data collected, the purposes of processing, categories of recipients, retention practices, significant risks, and available choices. Interfaces should avoid dark patterns such as preselected permissions, confusing language, or making refusal substantially more difficult than acceptance.

Consent is especially problematic in unequal relationships. Employees may feel unable to refuse workplace monitoring; students may believe that refusal will affect educational opportunities; patients may not understand secondary research uses; and users may have little practical choice when a dominant platform controls access to an essential service. In these situations, organisations should consider whether another lawful and ethically defensible basis exists, whether the activity is necessary, and whether an alternative service can be provided without coercion.

The principle of contextual integrity provides a valuable analytical lens. Information is appropriate when it flows in ways consistent with the social context in which it was generated. A person may share a photograph with friends without expecting it to train a facial-recognition system. A patient may disclose information to a clinician without expecting it to be sold to an advertiser. A user’s public post may be technically visible to anyone but still carry a contextual expectation against automated profiling. Ethical evaluation should therefore examine the norms and expectations surrounding the original disclosure, not only the visibility of the data.

Consent also has temporal limits. A person may agree to a service today but not to an unrelated future purpose. Organisations should reassess consent when purposes, recipients, model capabilities, or risks change. For children and other vulnerable groups, stronger safeguards, age-appropriate communication, and guardian or institutional protections may be required, while recognising that guardian permission does not eliminate the child’s own interests.

Provenance, Representation, and Data Quality

Data provenance is the documented history of a dataset: its sources, collection methods, transformations, labels, ownership claims, licences, access restrictions, and subsequent uses. Provenance enables an organisation to answer questions such as: Who collected the data? Was collection lawful and ethical? Were people informed? What preprocessing occurred? Which populations are missing? What restrictions travel with the data? Can a problematic record or source be removed?

A provenance record commonly includes a data inventory, source description, collection date, consent or legal basis, licence, labelling protocol, transformation history, known limitations, quality tests, and responsible owner. Dataset documentation and model documentation can make these matters visible to developers and reviewers. Provenance is not merely administrative. It supports reproducibility, auditability, incident response, intellectual-property analysis, and informed model evaluation.

Representativeness is similarly more complex than achieving a balanced sample. A dataset can reflect the distribution of a population but still encode historical discrimination, measurement error, or inappropriate labels. A facial dataset may contain many demographic categories but underrepresent lighting conditions or cultural practices. A hiring dataset may accurately mirror an organisation’s past workforce while reproducing exclusionary recruitment decisions. Representation must therefore be assessed in relation to the intended task, deployment context, affected populations, and consequences of error.

Important questions include:

1. Which populations are included, excluded, or systematically mismeasured?

2. Are labels reliable, and whose judgments produced them?

3. Does the training distribution match the deployment environment?

4. Are minority groups sufficiently represented for meaningful performance evaluation?

5. Could collection itself expose a vulnerable group to harm?

Data quality dimensions include accuracy, completeness, consistency, timeliness, validity, and relevance. Yet improving quality is not always ethically neutral. Removing unusual records may erase minority experiences. Imputing missing values may conceal structural exclusion. A responsible team records such decisions and evaluates their effects rather than treating cleaning as an objective technical step.

Privacy Across the AI Lifecycle

Privacy risks can enter at every stage. During collection, organisations may gather more information than necessary. During preparation, multiple datasets may be linked, creating new inferences. During training, a model may memorise rare records. During deployment, predictions may reveal sensitive attributes or enable continuous monitoring. During retention, old data may remain vulnerable despite no longer serving a legitimate purpose.

Data minimization requires collecting and retaining only what is reasonably necessary for a specified purpose. It is not simply deleting data after collection. It begins with problem formulation: can the objective be achieved with fewer variables, lower precision, shorter retention, local processing, or synthetic information? For example, a building-access system may need a yes/no authorization result rather than a permanent record of every movement. A service may need an age category rather than an exact birth date.

Minimization should be combined with purpose limitation, access control, retention schedules, secure deletion, and privacy impact assessment. Access should follow role-based or attribute-based principles, with logging and periodic review. Encryption protects data in storage and transit, but it does not solve inappropriate collection or excessive access. Security and privacy are related: a breach can expose personal data, while a perfectly secure system can still violate privacy through unjustified surveillance.

Anonymization and Its Limits

Anonymization aims to prevent identification of individuals, either directly or indirectly. Direct identifiers such as names and identity numbers are easy to remove, but quasi-identifiers—date of birth, location, occupation, or rare combinations of attributes—may permit re-identification when linked with external information. Anonymization is therefore a risk assessment, not a formatting operation. It must consider the data, likely adversaries, available auxiliary datasets, and future changes in technology.

Pseudonymization replaces identifiers with codes while retaining a means of re-linking. It reduces exposure in routine processing but remains personal data where re-identification is possible. Aggregation reports group-level statistics and can reduce individual-level risk, although small groups and repeated queries may still reveal information. Suppression, generalisation, perturbation, and controlled access are additional techniques.

A further concern is inference privacy. A system may not disclose a stored attribute yet infer it accurately from seemingly harmless variables. For example, mobility, purchasing, or language patterns may reveal health, political affiliation, or economic status. Privacy evaluation must therefore consider what a model can infer, not only what the database explicitly contains.

Privacy-Enhancing Technologies

Privacy-enhancing technologies (PETs) support useful analysis while reducing exposure of raw data. Their suitability depends on the application, threat model, accuracy requirements, and operational capacity.

• Differential privacy adds calibrated statistical noise so that the presence or absence of one individual has limited influence on an output. Its privacy budget makes repeated queries and accuracy trade-offs explicit.
• Federated learning trains models across distributed devices or organisations so that raw data can remain local. Model updates may still leak information and often require secure aggregation and additional protections.
• Secure multi-party computation allows parties to compute a function jointly without revealing their inputs, although computational and communication costs may be substantial.
• Homomorphic encryption permits certain computations on encrypted data, reducing exposure during processing but often imposing performance constraints.
• Synthetic data generates artificial records with statistical properties resembling source data. It can assist testing and sharing, but it may reproduce source bias or leak unusual records and should not automatically be treated as anonymous.
• Trusted execution environments isolate computation in protected hardware, but their security depends on implementation, configuration, and trust assumptions.

PETs are not ethical substitutes for purpose limitation or governance. A technically advanced privacy mechanism cannot justify an unnecessary or discriminatory use case. Teams should document the threat model, privacy guarantee, utility loss, residual risks, and operational controls.

Surveillance and the Power to Observe

AI changes surveillance by making observation scalable, continuous, and inferential. Cameras, location systems, keystroke monitoring, emotion-recognition claims, productivity scores, and predictive policing tools can transform occasional observation into persistent behavioural evaluation. The ethical issue is not only whether data is accurate, but whether people can live, work, learn, and participate without unreasonable monitoring.

A proportionality assessment asks whether the purpose is legitimate, whether surveillance is necessary to achieve it, whether a less intrusive alternative exists, and whether benefits justify the effects on autonomy and dignity. Notice alone is insufficient when individuals cannot realistically avoid observation. Organisations should define prohibited uses, limit secondary analytics, provide avenues for contesting decisions, and ensure that monitoring does not become a proxy for trust.

Surveillance can produce chilling effects: people may avoid lawful speech, medical support, association, or experimentation because they believe their behaviour is being recorded and interpreted. Errors also have unequal consequences. A false alert in a low-stakes marketing system differs from a false suspicion in employment, education, immigration, or policing. High-impact applications require stronger validation, human oversight, independent review, and meaningful remedies.

Intellectual Property and the Data Supply Chain

Intellectual property (IP) includes copyright, patents, trade secrets, trademarks, and rights associated with databases or distinctive creations. AI projects encounter IP at several layers: source materials, code and libraries, labels, trained parameters, model architecture, prompts, and outputs.

Copyright generally protects original expression, not abstract facts or ideas, but the legal treatment of datasets, compilations, annotations, and machine-learning activities varies by jurisdiction and licence. Public availability is not the same as permission for every form of copying, scraping, commercialisation, or redistribution. A responsible organisation records the source and licence of each material, checks attribution and share-alike requirements, respects access controls, and obtains legal advice when uncertainty is material.

Open-source software also carries obligations. Licences may require notices, disclosure of modifications, source availability, or restrictions on redistribution. Data licences may prohibit commercial use, require attribution, or restrict combination with other sources. A model trained on material under incompatible licences can create downstream distribution problems. Trade secrets introduce a different concern: an organisation must avoid using confidential information obtained without authorisation, even if a model makes it difficult to identify the original source.

Ownership of AI-generated outputs is legally unsettled in some jurisdictions and may depend on human contribution, contractual terms, and the nature of the output. Organisations should define ownership, permitted use, attribution, confidentiality, and review responsibilities in contracts and internal policies. They should also avoid representing generated content as wholly original when it substantially reproduces protected expression.

The ethical dimension extends beyond legal ownership. Data labelling and content moderation often rely on invisible or underpaid labour. Indigenous knowledge, community archives, and culturally significant materials may require collective consent and restrictions that are not captured by conventional individual copyright. Respectful data governance recognises contributors, communities, and provenance, not only the institution that stores the files.

A Lifecycle Governance Framework

A practical governance framework can be organised around six decision gates.

Gate 1: Justify the objective. Define the problem, expected benefit, affected stakeholders, foreseeable harms, and whether AI is necessary. Reject uses whose purpose is vague, disproportionate, or incompatible with rights.

Gate 2: Authorise acquisition. Record source, consent or other legal basis, licence, collection context, sensitivity, representativeness, and restrictions. Conduct a privacy or ethical impact assessment before large-scale collection.

Gate 3: Prepare responsibly. Document labelling, cleaning, linkage, feature engineering, anonymization, access controls, and decisions affecting minority groups. Preserve version history.

Gate 4: Evaluate before deployment. Test accuracy, calibration, subgroup performance, privacy leakage, memorization, robustness, and possible misuse. Include affected stakeholders and domain experts, not only data scientists.

Gate 5: Deploy with controls. Use least-privilege access, monitoring, human review where warranted, user notices, appeal mechanisms, retention rules, incident response, and change management.

Gate 6: Review and retire. Monitor distribution shift, new harms, complaints, legal developments, and changing consent expectations. Delete or archive data according to documented rules and securely retire models that no longer meet requirements.

This framework works best when responsibility is distributed but explicit. A data owner may approve business use, a privacy officer may assess risks, legal specialists may review licences, engineers may implement controls, and an ethics or governance committee may review high-impact applications. Clear escalation paths prevent “everyone is responsible” from becoming “no one is accountable.”

Illustrative Case Study: An AI-Based Employee Wellbeing Platform

Background

A large organisation considers deploying an AI platform that analyses email metadata, meeting schedules, badge records, and optional employee surveys to identify teams at risk of burnout. The vendor promises early intervention and claims that the system does not read email content.

Situation and Problem

Management wants weekly risk scores for team leaders. Employees are informed through a general privacy notice, but participation is not separately requested. The proposed data includes precise timestamps and location information. The vendor retains identifiable records for two years and may use aggregated data to improve its commercial product.

Analysis

The project has a plausible benefit, but several risks are present. Employees may not feel free to refuse, making consent questionable. Metadata can reveal medical appointments, union activity, religious observance, or personal relationships even without message content. The model may reproduce managerial or occupational stereotypes and may interpret reduced meeting attendance as disengagement rather than workload or accessibility needs. Team-level scores can stigmatise groups and encourage punitive monitoring. The vendor’s secondary use creates a purpose and IP concern, particularly if employees were not informed that their data would contribute to a commercial model.

Approach and Strategy

A responsible redesign begins by questioning whether individual-level prediction is necessary. The organisation could use voluntary, anonymous pulse surveys and aggregated workload indicators, with minimum group sizes that prevent re-identification. Precise location and badge data should be excluded unless a documented necessity is demonstrated. The organisation should conduct a data-protection and human-rights impact assessment, consult employees and representatives, and explain the system in accessible language.

If deployment proceeds, data should be minimised, retained briefly, separated from performance-management systems, and protected through role-based access. Results should support resources such as staffing or counselling rather than discipline. The vendor contract should prohibit secondary use, specify deletion, require audit access, address model updates, and clarify ownership and permitted use of derived models. Independent evaluation should test whether predictions differ unfairly across job roles, disability status, work arrangements, and demographic groups. Employees should have a channel to challenge records and decisions.

Outcome

The organisation rejects individual risk scores and adopts an aggregated planning tool. It uses the tool to identify workload concentration at the department level, publishes its methodology, limits retention, and reviews the programme every six months. The resulting system provides less granular prediction but better aligns with employee autonomy and proportionality.

Learning Points

The case demonstrates that the best ethical decision may be to change the objective rather than merely secure the original system. Metadata can be sensitive; consent is weakened by workplace power; aggregation does not automatically eliminate risk; and vendor contracts are central instruments of privacy and IP governance.

Challenges, Common Errors, and Professional Judgement

A common mistake is treating a privacy notice as a substitute for meaningful choice. Another is assuming that removing names creates anonymity. Teams also frequently neglect data acquired from vendors, presuming that contractual access proves ethical provenance. Benchmark datasets may be convenient but poorly matched to deployment populations. “More data” may increase accuracy on an average metric while worsening harms for underrepresented groups.

Organisations may over-rely on checklists. A completed impact assessment is valuable only if it changes design, procurement, or deployment decisions. Similarly, fairness and privacy cannot be assessed once and permanently certified: data distributions, threats, laws, and social expectations change. Excessive de-identification can reduce usefulness, while insufficient protection can expose people. Professional judgement requires making these trade-offs explicit and documenting why a chosen level of risk is acceptable.

Another challenge is collective harm. A model may not identify individuals but still classify a neighbourhood as risky, reduce services to a group, or exploit a community’s cultural material. Governance must therefore consider group privacy and social effects in addition to individual rights. Finally, organisations should prepare for incidents: unauthorised access, licence disputes, re-identification, model extraction, or evidence that data was collected deceptively. Response plans should include containment, notification where required, correction, deletion, remediation, and learning.

Current Developments and Best Practices

Recent AI development has intensified scrutiny of large-scale web scraping, foundation-model training, synthetic data, biometric identification, and generative systems that may reproduce protected or personal content. Regulatory approaches increasingly emphasise risk classification, transparency, documentation, human oversight, data governance, and accountability across the supply chain. Requirements differ by jurisdiction, so organisations operating internationally need a legal mapping process rather than assuming one global rule.

Generative AI introduces additional concerns: prompts may contain confidential information; retrieval systems may expose documents to unauthorised users; outputs may reproduce memorised personal or copyrighted material; and provenance may be difficult to establish. Organisations should define approved tools, prohibit entry of sensitive data into unmanaged services, log access appropriately, test for memorization and leakage, and require human review for high-impact outputs.

A mature programme commonly adopts the following practices:

• maintain a living data and model inventory;
• document provenance, licences, consent, transformations, and limitations;
• conduct privacy, ethical, and human-rights impact assessments proportionate to risk;
• collect the minimum data necessary and establish deletion triggers;
• use PETs according to a documented threat model;
• test representation, performance, and privacy across relevant groups;
• include affected stakeholders in design and review;
• build contractual controls into procurement and vendor management;
• provide notice, access, correction, explanation, and appeal mechanisms where appropriate;
• train staff in privacy, copyright, security, and responsible AI;
• audit continuously and preserve evidence for accountability.

Chapter Summary

Responsible AI depends on responsible data governance. Consent must be meaningful and context-sensitive, while provenance establishes where data came from, how it changed, and what restrictions apply. Representativeness requires attention to missing populations, biased labels, and deployment context rather than simple statistical balance. Privacy is supported by minimization, purpose limitation, access control, retention limits, and technologies such as differential privacy, federated learning, secure computation, and encryption. These technologies reduce risk but do not justify unnecessary collection or harmful purposes.

Surveillance requires proportionality because AI can make observation continuous and inferential. Intellectual-property analysis must cover source materials, datasets, software, labels, models, and outputs, with careful attention to licences, attribution, confidentiality, and community rights. A lifecycle governance framework links problem justification, authorised acquisition, responsible preparation, pre-deployment evaluation, controlled deployment, and ongoing review. The central professional obligation is to make data-related choices transparent, proportionate, contestable, and accountable to the people and communities affected.

Key Terms

Data ethics: Moral principles governing the collection, use, sharing, retention, and deletion of data.

Privacy: An individual’s or group’s interest in controlling information and participation in data-related decisions.

Consent: An informed, specific, voluntary indication of agreement to a defined data use.

Contextual integrity: The view that appropriate information flows depend on the norms of the social context.

Data provenance: The documented history and lineage of data sources, transformations, permissions, and uses.

Representativeness: The degree to which data reflects relevant populations and conditions of intended use.

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

Purpose limitation: Restricting data use to purposes identified and justified at collection.

Anonymization: Processing intended to prevent identification of individuals, including through indirect attributes.

Pseudonymization: Replacing identifiers with codes while retaining a possible method of re-identification.

Inference privacy: Protection against revealing sensitive information through predictions or correlations.

Differential privacy: A formal privacy approach limiting the influence of any one record on an output.

Federated learning: Distributed model training in which raw data remains at local sites or devices.

Privacy-enhancing technology: A technical method that supports data use while reducing privacy exposure.

Surveillance: Systematic observation, recording, or inference about people or groups.

Intellectual property: Legally recognised interests in creations, inventions, software, data arrangements, and brands.

Copyright: Protection for qualifying original expression and certain compilations under applicable law.

Trade secret: Confidential information protected through secrecy and reasonable protective measures.

Purpose creep: Expansion of data use beyond the purpose for which it was originally collected or expected.

Data governance: Policies, roles, controls, and processes that direct responsible data management.

Review Questions

Short Answer Questions

1. Why is legal compliance not sufficient for ethical data governance?

2. Distinguish anonymization from pseudonymization.

3. What information should a data-provenance record contain?

4. Explain why public availability does not automatically grant permission to use data for AI training.

5. What is the purpose of data minimization?

Long Answer Questions

1. Explain how consent, contextual integrity, and power relationships affect the legitimacy of AI data collection.

2. Discuss privacy risks across the AI lifecycle and propose controls for each stage.

3. Compare differential privacy, federated learning, secure multi-party computation, and synthetic data.

4. Analyse the intellectual-property issues that may arise in datasets, open-source software, trained models, and generated outputs.

5. Develop a lifecycle governance framework for a high-impact AI application.

Analytical / Application Questions

1. A hospital wants to share a de-identified patient dataset with a commercial AI company. Identify the provenance, consent, re-identification, licensing, and governance questions that should be resolved before sharing.

2. A university proposes using student Wi-Fi and learning-platform logs to predict students at risk of dropping out. Evaluate the proposal using minimization, proportionality, contextual integrity, representativeness, and appeal mechanisms.

3. A company discovers that its generative model sometimes reproduces passages from copyrighted books. Design an immediate response and a longer-term training-data and governance strategy.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes ethical data governance from legal compliance?

A. Ethical governance requires ignoring applicable law when social benefits are high.

B. Legal compliance is sufficient whenever users have clicked an agreement.

C. Ethical governance also considers fairness, context, proportionality, dignity, and foreseeable harm.

D. Ethical governance applies only to publicly available data.

Correct Answer: C. Ethical governance also considers fairness, context, proportionality, dignity, and foreseeable harm.

Question 2

A company removes names from a dataset but retains precise location, occupation, and date-of-birth information. What is the principal concern?

A. The data cannot be analysed statistically.

B. Quasi-identifiers may enable re-identification when combined with external information.

C. Removing names automatically makes the dataset synthetic.

D. The dataset has become intellectual property.

Correct Answer: B. Quasi-identifiers may enable re-identification when combined with external information.

Question 3

Which practice most directly implements data minimization?

A. Collecting every potentially useful variable for future projects.

B. Retaining raw records indefinitely for reproducibility.

C. Collecting only the information necessary for a specified purpose and deleting it when no longer needed.

D. Publishing all collected information to increase transparency.

Correct Answer: C. Collecting only the information necessary for a specified purpose and deleting it when no longer needed.

Question 4

What is a central feature of differential privacy?

A. It guarantees that no prediction can ever be made about an individual.

B. It limits the influence of any one person’s record on a released result.

C. It requires all data to be stored on a single secure server.

D. It removes the need for access controls.

Correct Answer: B. It limits the influence of any one person’s record on a released result.

Question 5

Why may employee consent to workplace monitoring be ethically weak?

A. Employees never have privacy interests at work.

B. Workplace data is always anonymous.

C. Unequal power may make refusal difficult or practically coercive.

D. Consent is unnecessary for all organisational data.

Correct Answer: C. Unequal power may make refusal difficult or practically coercive.

Question 6

Which item is most important for establishing data provenance?

A. The colour scheme used in a dashboard.

B. The dataset’s file size alone.

C. The source, collection context, transformations, permissions, and known limitations.

D. The model’s accuracy without reference to the data.

Correct Answer: C. The source, collection context, transformations, permissions, and known limitations.

Question 7

A model is trained using material found on public websites. Which conclusion is most defensible?

A. Public visibility automatically permits commercial AI training.

B. Copyright and licence conditions may still restrict copying, training, or redistribution.

C. All web material is in the public domain.

D. Attribution is never relevant to machine-learning data.

Correct Answer: B. Copyright and licence conditions may still restrict copying, training, or redistribution.

Question 8

Which redesign best reduces the risk in the employee wellbeing case?

A. Generate individual risk scores and send them to supervisors.

B. Retain detailed location histories for two years.

C. Use aggregated indicators, minimise collection, prohibit punitive use, and provide employee consultation.

D. Hide the system from employees to prevent gaming.

Correct Answer: C. Use aggregated indicators, minimise collection, prohibit punitive use, and provide employee consultation.

Question 9

Why can synthetic data still create privacy or fairness risks?

A. It is always identical to the original records.

B. It may reproduce source bias or reveal unusual patterns from the original data.

C. It cannot be used for testing.

D. It eliminates the need for provenance documentation.

Correct Answer: B. It may reproduce source bias or reveal unusual patterns from the original data.

Question 10

Which sequence best represents responsible AI data governance?

A. Deploy first, acquire data later, and document only after complaints.

B. Justify the objective, authorise acquisition, prepare responsibly, evaluate, deploy with controls, and review or retire.

C. Collect broadly, train quickly, and rely solely on model accuracy.

D. Anonymize once and permanently eliminate all governance duties.

Correct Answer: B. Justify the objective, authorise acquisition, prepare responsibly, evaluate, deploy with controls, and review or retire.