Law, Regulation, Standards, and Professional Duties

Introduction and Context

Artificial intelligence is governed not by one universal law but by a layered system of constitutional principles, legislation, regulatory guidance, technical standards, contractual duties, and professional ethics. An AI developer may therefore have to satisfy several requirements simultaneously: protect personal data, avoid unlawful discrimination, provide consumers with accurate information, respect intellectual property, ensure product safety, and allocate responsibility when an AI system causes harm.

This legal environment is increasingly risk-based. Rather than prohibiting all AI or regulating every system identically, contemporary frameworks distinguish between acceptable, restricted, high-risk, and prohibited uses. Risk-based regulation is attractive because it can focus institutional resources on systems capable of causing serious harm. It also creates difficulties: risk is context-dependent, classifications may be contested, and formal compliance cannot guarantee ethical outcomes.

This chapter examines major international and European frameworks, with particular attention to the EU Artificial Intelligence Act, the NIST AI Risk Management Framework, ISO/IEC 42001, the OECD AI Principles, UNESCO’s Recommendation on the Ethics of Artificial Intelligence, and professional codes. The objective is not to treat compliance as a checklist. Responsible innovation requires translating legal and normative duties into decisions about system design, procurement, deployment, monitoring, and retirement.

Learning Objectives

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

1. Explain risk-based regulation and distinguish legal rules, standards, guidance, and professional duties.

2. Compare the principal obligations created by data protection, anti-discrimination, consumer protection, intellectual-property, product-safety, and liability regimes.

3. Interpret the structure and risk categories of the EU AI Act.

4. Explain how NIST AI RMF and ISO/IEC 42001 support organisational governance.

5. Relate OECD, UNESCO, and professional principles to practical engineering decisions.

6. Analyse responsibility and propose a defensible governance approach for an AI deployment.

1. A Layered Model of AI Governance

1.1 Law, regulation, standards, and ethics

Law consists of binding rules adopted by a competent authority. Breach can result in penalties, invalid contracts, damages, or criminal sanctions. Regulation may refer either to law generally or to the work of public authorities that interpret, supervise, and enforce legal requirements. Standards are agreed specifications or guidance, often developed by bodies such as ISO, IEC, or national standards organisations. A standard may be voluntary, but it can become practically important when incorporated into a contract, cited by legislation, or used as evidence of reasonable practice.

Soft law includes principles, codes, recommendations, and regulatory guidance. It may not create a direct legal cause of action, but it can influence enforcement, procurement, professional discipline, and judicial assessments of reasonable conduct. Professional duties arise from codes of conduct, employment responsibilities, accreditation requirements, and the engineer’s obligation to avoid foreseeable harm.

These layers interact. For example, a data scientist may use an ISO management system to organise compliance with data-protection law, use NIST guidance to identify model risks, and follow a professional code when the law does not resolve whether deployment is socially responsible.

1.2 Risk-based regulation

A risk-based approach generally involves four steps:

1. Identify the activity and affected parties. The same model may have different consequences in advertising, education, employment, or medical diagnosis.

2. Assess likelihood and severity of harm. Harms may include physical injury, financial loss, privacy intrusion, discrimination, exclusion, misinformation, or loss of autonomy.

3. Apply controls proportionate to risk. Controls may include testing, human oversight, documentation, access restrictions, impact assessments, or prohibition.

4. Monitor and revise. AI systems change through retraining, data drift, altered use, and integration with other systems.

Risk is not simply a property of an algorithm. It is produced by the relationship between a system, its deployment context, human decisions, and affected communities. A highly accurate model can still be high risk if errors have severe consequences or if people cannot challenge decisions.

A central limitation is that risk categories may conceal distributional effects. A system with a low average error rate may systematically disadvantage a minority group. Responsible governance therefore examines both overall performance and its allocation across populations.

2. The EU Artificial Intelligence Act

The EU AI Act establishes a horizontal framework for AI placed on the European Union market or used in ways within its territorial scope. It applies according to the system’s function, role in the value chain, and risk. The Act is progressively applicable rather than becoming operative in every respect on one date; organisations must consult the current official implementation timetable and national guidance.

2.1 Risk categories

The Act’s structure includes four important categories.

Prohibited or unacceptable-risk practices are banned because they conflict with fundamental rights or create unacceptable manipulation, exploitation, or social-control risks. The precise list is statutory and should be checked directly rather than inferred from general descriptions. Examples include certain manipulative practices, specified forms of social scoring, and some biometric uses.

High-risk systems are subject to extensive requirements. They include certain AI systems used as safety components of regulated products and systems listed in sensitive areas such as employment, education, essential services, law enforcement, migration, and justice. Classification depends on the Act’s conditions; a system is not automatically high risk merely because it uses machine learning.

Limited-risk systems may trigger transparency duties. For example, people may need to know when they are interacting with an AI system, and synthetic or manipulated content may require disclosure or machine-readable marking in specified circumstances.

Minimal- or low-risk systems are generally not subject to the Act’s strongest mandatory controls, although voluntary codes and other laws still apply. A low-risk classification does not remove duties under privacy, consumer, equality, safety, or intellectual-property law.

2.2 Duties relevant to high-risk systems

High-risk providers are expected to establish risk-management processes, use appropriate data-governance practices, prepare technical documentation, maintain records, provide instructions, support human oversight, achieve appropriate levels of accuracy, robustness, and cybersecurity, and operate a post-market monitoring process. Quality management and conformity-assessment obligations may also apply. Deployers have their own duties, including using systems according to instructions, assigning competent human oversight, monitoring operation, and responding to incidents.

The Act also distinguishes providers, deployers, importers, distributors, authorised representatives, and affected persons. Responsibility can shift when an organisation substantially modifies a system, changes its intended purpose, or places its own name on it. Procurement contracts should therefore specify intended use, data responsibilities, incident reporting, audit rights, model-change notification, and exit arrangements.

2.3 Interpreting the Act in practice

A useful compliance analysis asks:

• What is the system’s intended purpose?
• Who is the provider and who is the deployer?
• Does the system fall within a prohibited practice or a listed high-risk use?
• What other legislation applies independently?
• What evidence demonstrates risk management, data quality, oversight, and monitoring?
• What information and remedies are available to affected people?

The Act does not replace human-rights analysis. A formally compliant system may still be challenged under privacy, equality, employment, consumer, or constitutional law.

3. Data Protection and Privacy

Data protection law regulates the processing of personal data, not merely data storage. Under the EU General Data Protection Regulation (GDPR), organisations need a lawful basis for processing, must observe purpose limitation and data minimisation, and must provide transparency, security, retention controls, and mechanisms for exercising data-subject rights. Special categories of personal data, such as health or biometric information used for uniquely identifying a person, receive stronger protection under specified conditions.

AI creates distinctive data-protection challenges. Training data may have been collected for another purpose; personal information may be reproduced in outputs; a model may enable re-identification; and inferred attributes can be sensitive even when they were not explicitly collected. Organisations should document data provenance, lawful basis, retention, access controls, and deletion or correction procedures.

Automated decision-making rules may apply where a person is subject to a decision based solely on automated processing that produces legal or similarly significant effects. The precise legal analysis depends on the decision, the degree of human involvement, applicable exceptions, and safeguards. A nominal human click is not necessarily meaningful human review. A reviewer should have authority, relevant information, sufficient time, and the ability to depart from the model.

A data protection impact assessment is appropriate, and in some circumstances required, when processing is likely to create high risks to individuals. It should describe the processing, assess necessity and proportionality, identify risks, and specify measures to address them. Privacy-enhancing techniques—such as aggregation, access limitation, pseudonymisation, secure computation, or differential privacy—can reduce exposure, but they do not automatically make processing lawful.

4. Discrimination and Equality

Anti-discrimination law may prohibit unequal treatment based on protected characteristics, including race, sex, disability, age, religion, or other characteristics defined by the applicable jurisdiction. Discrimination may be direct, indirect, structural, or intersectional. AI can reproduce historical inequality, use proxies for protected attributes, create unequal error rates, or make a formally neutral rule that has disproportionate effects.

A common evaluation mistake is to search for one universal fairness metric. Fairness criteria can conflict. For example, equal selection rates, equal false-positive rates, and equal predictive accuracy across groups may not all be simultaneously achievable when groups have different base rates. Consequently, technical testing must be connected to the legal context, institutional purpose, and affected groups.

A defensible process includes:

• identifying legally and socially relevant groups;
• examining representation and label quality;
• testing performance and error patterns by group;
• evaluating accessibility and reasonable accommodation;
• consulting affected stakeholders;
• documenting trade-offs and decisions;
• providing notice, explanation, appeal, and correction mechanisms.

Fairness testing is not a substitute for legal advice or institutional reform. Sometimes the correct response is to change the decision process, remove an AI component, or stop deployment.

5. Consumer Protection and Transparency

Consumer-protection law generally prohibits misleading claims, unfair commercial practices, hidden material information, and unsafe or defective products. AI providers should not claim that a system is unbiased, autonomous, accurate, or secure without evidence and appropriate qualification. Marketing language can create legal exposure even where the technical model itself performs as described.

Users should understand when they are interacting with AI, what the system can and cannot do, whether content is synthetic, how personal data are used, and how to obtain human assistance. Explanations should be usable rather than merely technically impressive. A statement such as “the model identified risk” does not tell an affected person what information mattered or how to challenge the outcome.

Dark patterns and persuasive personalisation create additional concerns. A system may exploit vulnerability by adapting messages to children, people in financial distress, or individuals with limited digital literacy. Responsible design uses clear choices, avoids coercion, and preserves the possibility of refusal.

6. Intellectual Property

AI development may implicate copyright, database rights, trademarks, patents, trade secrets, and rights of publicity. Key questions include whether training or retrieval copies are lawful, whether outputs reproduce protected expression, who owns or licenses generated material, and whether confidential information has been disclosed to a model service.

The answers differ by jurisdiction and by the facts of the use. A model’s output is not automatically free of infringement merely because it was generated by software. Conversely, not every resemblance proves infringement. Organisations should maintain data and licence inventories, respect contractual restrictions, use approved repositories, prevent confidential prompts from entering external services, and review outputs before commercial publication.

Open-source and open-weight models require careful licence analysis. “Open” may refer to source code, weights, data, or documentation, and each licence may impose different attribution, notice, redistribution, or usage conditions. Procurement teams should record model versions and dependencies so that obligations remain traceable.

7. Product Safety and Liability

When AI controls or influences a physical product, safety becomes central. Examples include vehicles, medical devices, industrial robots, and consumer appliances. Safety engineering requires hazard identification, foreseeable-misuse analysis, validation, fail-safe behaviour, cybersecurity, human-factors evaluation, and post-market surveillance.

Liability may arise through negligence, contract, consumer-protection law, product-liability regimes, professional negligence, breach of statutory duty, or specific AI legislation. The appropriate defendant may be a developer, manufacturer, deployer, operator, integrator, or service provider. Complex supply chains make contractual allocation important, but a contract cannot always eliminate duties owed to injured third parties.

AI complicates causation and evidence. A claimant may need to show that a defect or unreasonable conduct caused harm, while the organisation controls logs, model versions, and design records. Good governance therefore includes tamper-resistant logging, version control, incident investigation, preservation of evidence, and clear escalation responsibilities.

Professional negligence is assessed against the standard of a reasonably competent practitioner in the circumstances. Following a recognised standard can support a defence, but certification is not immunity. Conversely, departure from a standard is not automatically unlawful; the surrounding facts matter.

8. NIST AI RMF and ISO/IEC 42001

8.1 NIST AI Risk Management Framework

The NIST AI RMF is a voluntary framework organised around four functions:

• Govern: establish policies, accountability, culture, legal and ethical responsibilities, and stakeholder engagement.
• Map: understand context, intended purpose, affected groups, risks, benefits, and limitations.
• Measure: test and monitor validity, reliability, safety, security, transparency, explainability, privacy, and fairness.
• Manage: prioritise risks, apply controls, document residual risk, and respond to incidents.

The framework is flexible rather than a certification scheme. Its value lies in connecting sociotechnical context to measurement and action. A team should not measure only model accuracy; it should also measure operational failures, complaints, subgroup impacts, security events, and whether human oversight functions in practice.

8.2 ISO/IEC 42001

ISO/IEC 42001 specifies requirements for an AI management system. It uses a management-system logic: define organisational context and leadership, plan for risks and opportunities, provide resources and competence, control operational processes, evaluate performance, and pursue continual improvement. It can help an organisation assign responsibilities, maintain documented information, audit processes, and implement corrective action.

NIST AI RMF is primarily a practical risk-management framework, while ISO/IEC 42001 is a formal management-system standard that may support certification through an appropriately accredited process. They can be integrated: NIST functions can inform operational risk activities, while ISO/IEC 42001 supplies organisational governance, audit, and improvement structures.

Neither framework determines whether a particular use is morally acceptable or legally permitted. They support governance; they do not replace statutes, rights analysis, or professional judgement.

9. OECD, UNESCO, and Professional Principles

The OECD AI Principles emphasise inclusive growth and sustainable development, human-centred values and fairness, transparency and explainability, robustness, security and safety, and accountability. They also stress international cooperation and responsible stewardship across the AI lifecycle.

UNESCO’s Recommendation on the Ethics of Artificial Intelligence adopts a human-rights-oriented approach. It highlights proportionality, safety and security, fairness and non-discrimination, sustainability, privacy, human oversight, transparency, explainability, responsibility, and democratic governance. It gives particular attention to vulnerable groups, environmental effects, education, cultural diversity, and the global distribution of benefits and risks.

Professional codes translate broad values into role responsibilities. Codes from computing and engineering bodies commonly require practitioners to contribute to public welfare, avoid harm, be honest about system limits, protect privacy, maintain competence, respect confidentiality, and disclose conflicts of interest. A professional may have a duty to escalate concerns even when a manager prefers rapid deployment.

These instruments are useful for resolving gaps. Law may not specify how much testing is enough, while a professional code may require candour and competence. A standard may organise evidence, while a human-rights principle may determine whether the objective itself is legitimate.

10. An Integrated Compliance and Responsibility Process

A practical lifecycle process can be organised as follows.

Stage 1: Define purpose and boundaries. State what the system will do, what it will not do, who may be affected, and what decisions remain human.

Stage 2: Classify legal and social risk. Identify relevant AI Act status, privacy obligations, equality concerns, consumer duties, IP constraints, safety requirements, and sector rules.

Stage 3: Assign roles. Name the provider, deployer, product owner, data controller or processor where relevant, safety lead, privacy lead, and incident owner.

Stage 4: Assess impacts. Conduct proportionate privacy, equality, safety, human-rights, and environmental assessments. Consult affected stakeholders rather than relying only on internal assumptions.

Stage 5: Design controls. Use data governance, access control, testing, human oversight, documentation, secure development, content labelling, appeal processes, and fallback procedures.

Stage 6: Validate before release. Test representative scenarios, subgroup performance, foreseeable misuse, adversarial behaviour, accessibility, and operational integration.

Stage 7: Deploy with monitoring. Track drift, incidents, complaints, overrides, disparities, security events, and changes in use.

Stage 8: Review and retire. Reassess after material changes and preserve records needed for audit, investigation, and legal rights. Retire the system when risks cannot be controlled or its purpose is no longer justified.

11. Illustrative Case: Automated Recruitment Screening

Background

A fictional company, Northstar Services, proposes an AI tool to rank applicants for customer-support jobs. It is trained on ten years of historical hiring decisions and produces a shortlist for recruiters.

Problem

Historical hiring data show that applicants from some neighbourhoods and applicants with career interruptions were selected less often. The model also uses employment history, education, writing samples, and inferred commuting convenience. Managers want rapid deployment and describe the tool as objective.

Analysis

The system is used in employment, a context requiring careful assessment under the EU AI Act and equality law. Historical labels may encode past discrimination. Commuting convenience may act as a proxy for socioeconomic status or disability. A high average prediction accuracy would not resolve unequal exclusion. The company must also determine its data-protection role, lawful basis, transparency duties, retention period, and procedures for applicant rights. Marketing the tool as “bias-free” could create consumer or employment-law concerns.

Approach

Northstar should first define whether the tool assists or determines recruitment. It should remove unjustified proxy features, assess data representativeness, conduct subgroup testing, and involve employment-law and privacy specialists. Recruiters must receive meaningful training and authority to reject the recommendation. Applicants should receive appropriate information and a route for human review. The organisation should document intended purpose, model limitations, test results, version history, incidents, and reasons for material changes. A pilot with clear stop criteria should precede wider deployment.

Outcome

Testing reveals that the model disadvantages applicants with career interruptions, including people returning from caregiving leave. Northstar abandons the original ranking objective and uses a narrower tool to organise applications against transparent, job-related criteria. It introduces structured human review and audits outcomes quarterly.

Learning Points

The case demonstrates that legal classification begins with purpose and context, not model architecture. Historical data are not neutral evidence. Human oversight must be substantive, and the responsible decision may be redesign or non-deployment rather than technical correction.

12. Contemporary Developments and Limitations

AI regulation is evolving rapidly. Generative AI has increased attention to synthetic content, copyright, foundation-model responsibilities, model evaluations, and concentration of infrastructure. Regulators are also examining biometric identification, workplace surveillance, children’s safety, environmental costs, cybersecurity, and cross-border data flows.

Several limitations remain. Jurisdictions differ in definitions, enforcement capacity, remedies, and treatment of innovation. Small organisations may struggle with documentation and specialist assessment. Audits can become symbolic if they lack independence or access to relevant data. Standards may be expensive or difficult to interpret, and certification can encourage confidence without proving real-world fairness. Risk assessments can also marginalise communities if affected people are not involved.

Best practice is therefore adaptive and evidence-based: use current legal advice, maintain a regulatory register, assign accountable leadership, document decisions, engage affected stakeholders, test sociotechnical outcomes, preserve auditability, and provide accessible remedies. Responsible innovation means asking not only whether a system can be built and deployed, but whether its purpose is justified, its benefits are fairly distributed, and its harms can be effectively prevented and repaired.

Chapter Summary

AI governance operates through interacting layers of law, regulation, standards, soft law, contracts, and professional duties. Risk-based regulation prioritises controls according to context, likelihood, and severity, but must account for unequal impacts and fundamental rights. The EU AI Act classifies practices by risk and assigns duties to providers, deployers, and other actors. Its requirements coexist with data-protection, equality, consumer-protection, intellectual-property, safety, and liability rules.

NIST AI RMF organises practical work through Govern, Map, Measure, and Manage. ISO/IEC 42001 provides a management-system structure for accountable and continually improving AI governance. OECD and UNESCO principles supply internationally influential values, while professional codes require competence, honesty, privacy protection, public welfare, and escalation of serious concerns. Effective practice combines classification, impact assessment, role allocation, technical and organisational controls, meaningful human oversight, monitoring, documentation, and accessible remedies.

Key Terms

Risk-based regulation: Regulation that applies controls in proportion to the likelihood and severity of potential harm.

Provider: An actor that develops an AI system or has it developed and places it on the market or into service under its name.

Deployer: An organisation or person using an AI system under its authority.

High-risk AI system: A system subject to enhanced legal requirements because of its function, product role, or deployment context.

Data protection: Legal governance of the processing of personal data.

Purpose limitation: Using personal data only for specified, explicit, and legitimate purposes.

Automated decision-making: Decision-making performed wholly or substantially through automated processing.

Meaningful human oversight: Human supervision with competence, information, time, and authority to intervene.

Discrimination: Unjustified unequal treatment or impact connected to protected or relevant characteristics.

Consumer protection: Rules preventing deception, unfair practices, and unsafe or defective products.

Intellectual property: Legal rights protecting creations, inventions, brands, databases, and confidential information.

Product liability: Responsibility for harm caused by defective or unsafe products.

NIST AI RMF: A voluntary framework organised around Govern, Map, Measure, and Manage.

ISO/IEC 42001: An international standard specifying requirements for an AI management system.

Technical documentation: Records explaining a system’s purpose, design, data, performance, risks, and controls.

Post-market monitoring: Ongoing observation of an AI system after deployment.

Soft law: Non-binding principles, recommendations, guidance, or codes that influence conduct and enforcement.

Professional duty: An obligation arising from competence, ethics, professional codes, employment, or accepted practice.

Review Questions

Short Answer Questions

1. What is meant by risk-based regulation?

2. Distinguish a provider from a deployer under the EU AI Act.

3. Name the four functions of the NIST AI RMF.

4. Why is meaningful human oversight more than a human clicking an approval button?

5. State two ways in which historical training data can create discrimination.

Long Answer Questions

1. Compare the EU AI Act, NIST AI RMF, and ISO/IEC 42001 in purpose, legal status, and practical use.

2. Explain how data protection and intellectual-property issues can arise during AI model development.

3. Discuss the relationship between technical fairness metrics and legal discrimination analysis.

4. Describe an end-to-end governance process for deploying a high-risk AI system.

5. Explain why professional codes remain important when detailed legislation exists.

Analytical / Application Questions

1. A university uses an AI system to rank scholarship applicants using grades, personal statements, postcode, and predicted future income. Identify the main legal and ethical risks and propose a governance process before deployment.

2. A retailer claims that its generative-AI customer service agent is “always accurate and unbiased.” Analyse the possible consumer-protection, discrimination, privacy, and professional-duty issues.

3. An industrial robot controlled by an adaptive AI model injures a worker after a software update. Identify the evidence, actors, controls, and liability questions that should be investigated.

Multiple Choice Questions with Answers

Question 1

Which NIST AI RMF function establishes organisational policies, accountability, and responsibilities?

A. Govern

B. Map

C. Measure

D. Manage

Correct Answer: A. Govern

Question 2

Why is a low average error rate insufficient to establish that an AI system is fair?

A. Accuracy is never relevant to AI systems

B. Errors may be distributed unequally across groups

C. Fairness requires removing all human involvement

D. Average error rates are illegal in every jurisdiction

Correct Answer: B. Errors may be distributed unequally across groups

Question 3

Which statement best describes ISO/IEC 42001?

A. A criminal statute governing biometric systems

B. A mandatory algorithm used to calculate fairness

C. A standard for an AI management system

D. A database-licensing regime

Correct Answer: C. A standard for an AI management system

Question 4

What is a central feature of risk-based regulation?

A. Every AI system receives identical controls

B. Only open-source systems are regulated

C. Controls are proportionate to context and potential harm

D. Technical performance replaces legal analysis

Correct Answer: C. Controls are proportionate to context and potential harm

Question 5

Which practice best demonstrates meaningful human oversight?

A. A reviewer automatically accepts every model recommendation

B. A reviewer has authority, information, competence, and time to intervene

C. A human is named in the system documentation but never reviews outputs

D. The organisation removes all audit logs

Correct Answer: B. A reviewer has authority, information, competence, and time to intervene

Question 6

Which issue is most directly associated with purpose limitation in data protection?

A. Using collected data for an unrelated objective without appropriate justification

B. Selecting a model with too few parameters

C. Changing a product logo

D. Using a longer software password

Correct Answer: A. Using collected data for an unrelated objective without appropriate justification

Question 7

Under the EU AI Act’s general structure, which category receives the strongest mandatory controls short of prohibited practices?

A. Minimal-risk systems

B. High-risk systems

C. Entertainment systems

D. Unclassified systems

Correct Answer: B. High-risk systems

Question 8

Why should an organisation maintain model-version and dependency records?

A. To guarantee that outputs are always creative

B. To support traceability, auditing, and incident investigation

C. To eliminate the need for human oversight

D. To make all intellectual-property rights expire

Correct Answer: B. To support traceability, auditing, and incident investigation

Question 9

Which statement about professional certification is most accurate?

A. Certification automatically eliminates liability

B. Certification proves that a system is morally acceptable

C. Certification can support evidence of process but does not replace legal duties

D. Certification makes privacy law inapplicable

Correct Answer: C. Certification can support evidence of process but does not replace legal duties

Question 10

In the recruitment case, why was historical hiring data problematic?

A. Historical data can encode previous discriminatory practices

B. Historical data can never be used for any purpose

C. Historical data automatically creates copyright infringement

D. Historical data prevents all model monitoring

Correct Answer: A. Historical data can encode previous discriminatory practices