AI, Automated Decision Systems and Responsible Innovation

Introduction and Context

Artificial intelligence (AI) has moved from specialist research laboratories into recruitment, credit, insurance, healthcare, policing, education, employment, public administration and consumer services. Many systems do not merely recommend content; they classify people, rank opportunities, allocate resources, detect suspected fraud, determine eligibility and generate documents that influence consequential decisions.

An automated decision system (ADS) is a computational system that uses data, rules, statistical models or machine-learning models to support or make decisions affecting individuals or organisations. Automation may be complete, as in an automatically rejected loan application, or partial, as in a system that ranks job applicants for human review. The legal and ethical significance of an ADS depends less on whether it is labelled “AI” than on its purpose, context, scale, affected rights and degree of human influence.

Responsible innovation treats compliance as more than a final legal check. It integrates law, ethics, information security, data governance, human rights, engineering assurance and organisational accountability throughout the system life cycle. This is especially important because AI risks can arise from design choices, training data, deployment context, user behaviour, vendor relationships and changes in the surrounding environment.

Learning Objectives

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

1. Distinguish AI systems, automated decision systems and generative-AI systems.
2. Explain algorithmic accountability, bias, fairness, transparency and explainability.
3. Assess privacy, intellectual-property, security and human-oversight risks.
4. Apply governance frameworks such as the NIST AI Risk Management Framework and risk-based regulatory approaches.
5. Select appropriate technical, organisational and procedural controls across an AI life cycle.
6. Analyse a consequential AI deployment and recommend a defensible governance approach.
7. Evaluate contemporary legal developments, including the EU AI Act and data-protection obligations.

1. From Algorithms to Sociotechnical Systems

An algorithm is a defined procedure for transforming inputs into outputs. AI systems may use symbolic rules, statistical learning, neural networks, reinforcement learning, retrieval systems or combinations of these techniques. A machine-learning model learns relationships from data rather than being programmed with every decision rule. A generative-AI model produces new text, images, audio, video or code in response to an input or prompt.

The phrase sociotechnical system is important. An AI model is only one component of a wider arrangement involving:

• data collection and labelling;
• software, infrastructure and interfaces;
• organisational objectives and performance incentives;
• human reviewers and affected persons;
• procurement contracts and third-party providers;
• legal rules and institutional procedures; and
• the social environment in which outputs are interpreted.

Consequently, a technically accurate model may still produce an unlawful or harmful outcome. For example, a model may predict employee attrition accurately while relying on proxies for disability, pregnancy or caring responsibilities. A facial-recognition system may have strong average accuracy but materially higher error rates for some demographic groups. A generative system may produce fluent but fabricated legal authorities.

An ADS should therefore be assessed at three levels:

2. Core Governance Principles

2.1 Accountability

Accountability means that a responsible person or organisation can be identified, decisions can be reviewed, evidence can be produced and remedies can be provided. “The algorithm decided” is not an adequate allocation of responsibility. Accountability should be assigned across the AI supply chain: deployer, provider, data controller, processor, system integrator, professional user and senior management.

A useful accountability structure records:

• the purpose and lawful basis of processing;
• the system owner and accountable executive;
• the model developer and version;
• data sources and known limitations;
• validation results and approval conditions;
• operational users and escalation routes;
• incidents, complaints and corrective actions; and
• conditions requiring suspension or withdrawal.

2.2 Legality, necessity and proportionality

Before asking whether a system can be built, an organisation should ask whether it should be used. A legitimate objective does not automatically justify every means of achieving it. A defensible assessment considers whether the objective is lawful, whether AI is necessary, whether a less intrusive method is available, and whether expected benefits are proportionate to risks.

Data-protection law is particularly relevant. Under the GDPR, organisations must address principles including lawfulness, fairness and transparency; purpose limitation; data minimisation; accuracy; storage limitation; integrity and confidentiality; and accountability. Article 22 addresses certain decisions based solely on automated processing that produce legal or similarly significant effects, subject to specified conditions and safeguards. The precise application depends on facts, jurisdiction and other legal instruments.

2.3 Human autonomy and oversight

Human oversight is meaningful only when the human has sufficient competence, authority, time, information and independence to intervene. A nominal “human in the loop” who approves nearly every model output without scrutiny does not provide robust oversight. Relevant design choices include:

• requiring review for high-impact or low-confidence cases;
• displaying reasons, uncertainty and relevant evidence;
• preventing automation bias through training and interface design;
• permitting override, correction and escalation;
• separating system operators from final decision-makers where conflicts exist; and
• recording the human rationale and action.

Human review cannot cure every problem. If reviewers receive biased data, face excessive workloads or lack authority to reject recommendations, the system remains effectively automated.

2.4 Safety, security and resilience

AI assurance includes ordinary cybersecurity as well as AI-specific threats. Risks include data poisoning, adversarial examples, model extraction, prompt injection, insecure plug-ins, unauthorised training-data disclosure and supply-chain compromise. Resilience requires tested fallbacks, incident response, access management, logging, backup processes and safe decommissioning.

2.5 Fairness and non-discrimination

Fairness is not one universal mathematical property. It is a normative and contextual judgement informed by law, institutional purpose and affected communities. Relevant questions include:

• Are protected groups excluded or disadvantaged?
• Are errors distributed unequally?
• Does the system reproduce historical discrimination?
• Are apparently neutral variables functioning as proxies?
• Is a different error trade-off justified by the application?
• Can affected people challenge inaccurate data or decisions?

For a binary classifier, a confusion matrix distinguishes true positives, false positives, true negatives and false negatives. Common measures include:

where TPR is the true-positive rate, TP is true positives and FN is false negatives. The false-positive rate is:

where FP is false positives and TN is true negatives. These measures are conditional rates and require a clearly defined positive outcome, reference population and decision threshold.

Equal opportunity generally seeks comparable true-positive rates across protected groups. Equalised odds seeks comparable true-positive and false-positive rates. Demographic parity compares the proportion receiving a positive outcome. Calibration asks whether predicted probabilities have comparable meaning across groups. These conditions can conflict, particularly where groups have different base rates. Therefore, a fairness assessment must explain the chosen criterion rather than claiming that one metric proves fairness.

2.6 Transparency and explainability

Transparency concerns what is disclosed about a system: its purpose, operator, data categories, limitations, decision role and governance. Explainability concerns how an output or behaviour is made intelligible. A system can be transparent without revealing every model parameter, and a technically interpretable model can still be used opaquely within an unfair process.

Explanations should be appropriate to the audience. A data scientist may need feature-importance analysis and validation statistics; an affected person may need a clear statement of the material factors, relevant data, decision-maker, correction route and appeal process. Post-hoc explanation tools can be useful, but they may be unstable, approximate or misleading. An explanation generated after a decision is not necessarily the actual causal basis of that decision.

3. Bias Across the AI Life Cycle

Bias can enter at every stage:

1. Problem formulation: the objective may encode an unjust or incomplete conception of success.
2. Sampling: some populations may be under-represented or systematically missing.
3. Measurement: labels may reflect subjective judgements or unequal observation.
4. Historical data: past institutional discrimination may be learned by the model.
5. Feature selection: proxies such as postcode, language or purchasing history may correlate with protected characteristics.
6. Modelling: optimisation may favour aggregate accuracy over minority performance.
7. Deployment: users may apply the system outside its validated context.
8. Feedback: predictions may alter behaviour, thereby changing future data.
9. Monitoring: performance may deteriorate because populations, policies or environments change.

Bias mitigation can occur before, during or after model training. Pre-processing may improve sampling or labels; in-processing may impose constraints during optimisation; post-processing may alter thresholds or outcomes. Each intervention has trade-offs. Removing a protected attribute does not remove discrimination if correlated proxies remain. Conversely, collecting protected-attribute data may be necessary for auditing, even where it should not be used for prediction.

A defensible assessment combines quantitative testing with qualitative inquiry. Disaggregated performance tables, error analysis, interviews, accessibility testing and consultation with affected communities can reveal harms that aggregate metrics conceal.

4. Privacy and Data Governance

AI systems intensify familiar privacy risks because they can infer sensitive attributes, combine datasets and retain information in models or logs. Privacy risks include excessive collection, function creep, re-identification, unlawful secondary use, insecure prompts, memorisation and disclosure through model outputs.

A practical AI data-governance programme should establish:

• data provenance and rights to use;
• purpose, lawful basis and retention period;
• data-quality and representativeness checks;
• access controls and segregation of sensitive data;
• deletion, correction and subject-rights procedures;
• rules for prompts, logs, embeddings and fine-tuning data;
• privacy-enhancing techniques where proportionate; and
• restrictions on transferring data to external providers.

De-identification reduces risk but does not guarantee anonymity. Pseudonymised data can remain personal data where re-identification is reasonably possible. Privacy-enhancing methods such as aggregation, access limitation, differential privacy and secure computation may help, but each has implementation costs and may reduce utility.

A privacy impact assessment should be integrated with an AI impact assessment rather than treated as an unrelated document. It should identify affected groups, likely inferences, data flows, residual risk and safeguards.

5. Intellectual Property and Generative AI

Generative AI creates difficult intellectual-property questions. Organisations must distinguish at least four issues:

1. Input rights: Does the organisation have permission to submit documents, code, personal data or confidential material to the provider?
2. Training rights: What materials were used to train or fine-tune the model, and what licences or legal exceptions apply?
3. Output rights: Is the output sufficiently original and legally protectable in the relevant jurisdiction?
4. Infringement and similarity: Does an output reproduce protected material or imitate a distinctive style in a legally actionable manner?

The legal position varies across jurisdictions and continues to develop. Copyright protection commonly depends on human authorship and other statutory requirements; an AI-generated output is not automatically owned by the person who entered a prompt. Businesses should not assume that provider terms transfer all rights or remove all liability.

Enterprise controls should include approved tools, confidential-data restrictions, source verification, licence review, code scanning, records of material human contribution and contractual allocation of responsibility. Generated content must be checked for hallucinations, fabricated citations, insecure code, discriminatory language and confidential-data leakage.

6. Generative-AI-Specific Risks

Generative systems introduce risks beyond conventional classification:

• Hallucination: plausible but false content presented with confidence.
• Prompt injection: instructions embedded in retrieved or supplied content manipulate system behaviour.
• Data leakage: sensitive information appears in prompts, outputs, logs or connected tools.
• Automation bias: users over-trust fluent outputs.
• Model collapse and quality degradation: synthetic data can contaminate later training.
• Abuse at scale: phishing, fraud, impersonation, disinformation and malicious code generation.
• Unclear provenance: users cannot determine whether content is authentic or licensed.
• Agentic overreach: a system connected to tools may take actions beyond the user’s intended authority.

A robust deployment uses least privilege, retrieval boundaries, output filtering, content provenance where feasible, human approval for consequential actions, adversarial testing and continuous monitoring. A chatbot should not be permitted to issue refunds, alter medical records or send legal notices merely because it can call an application programming interface.

7. Risk-Based Regulation and Standards

Contemporary regulation increasingly uses a risk-based model rather than treating all AI applications identically. The European Union’s AI Act, Regulation (EU) 2024/1689, entered into force in 2024 and establishes obligations that phase in over time. It prohibits specified unacceptable practices, imposes obligations on high-risk systems, introduces transparency duties for certain systems and establishes obligations for providers of general-purpose AI models. The Act interacts with other EU law, including data protection, product safety and sectoral regulation. Organisations must assess whether they are providers, deployers, importers, distributors or other actors and should consult the current legal text and competent authorities.

The GDPR remains relevant where personal data is processed, including by AI systems. Other jurisdictions have adopted or proposed sectoral, state-level or national measures. Regulatory obligations may concern impact assessments, documentation, notices, data governance, testing, monitoring, incident reporting, human oversight and record keeping.

The NIST AI Risk Management Framework is a voluntary framework organised around Govern, Map, Measure and Manage. It supports risk identification and treatment but does not itself create legal compliance. ISO/IEC 42001 provides requirements for an AI management system, while ISO/IEC 23894 addresses AI risk management. The OECD AI Principles emphasise inclusive growth, human-centred values, transparency, robustness, security, safety and accountability. These instruments can complement, but cannot replace, binding law.

Figure: AI governance is a continuous life-cycle process rather than a one-time approval.

8. An Operational Governance Model

A mature organisation can implement the following control sequence:

Stage 1: Register and classify

Create an inventory of AI systems, owners, providers, data categories, affected populations and business processes. Classify systems by impact, autonomy, scale and legal sensitivity.

Stage 2: Assess before procurement or development

Conduct legal, privacy, security, equality, safety and human-rights assessments. Define prohibited uses, performance requirements, escalation thresholds and acceptable residual risk.

Stage 3: Govern data and suppliers

Verify provenance, licences, quality and representativeness. Contracts should address audit rights, subcontractors, incident notification, data retention, security, model changes, service levels and exit assistance.

Stage 4: Test and validate

Use representative test data, subgroup analysis, robustness tests, red teaming and scenario-based evaluation. Validate not only model accuracy but also workflow outcomes, accessibility and human factors.

Stage 5: Approve with conditions

An approval should specify intended use, prohibited use, operating thresholds, required human review, monitoring indicators and expiry or reapproval dates.

Stage 6: Monitor and remedy

Track drift, error rates, subgroup outcomes, complaints, overrides, security events and near misses. Provide accessible correction and appeal channels. Suspend the system when risks exceed approved limits.

Stage 7: Retire responsibly

Preserve records required for accountability, revoke access, delete or return data where required, assess downstream dependencies and communicate the withdrawal.

Worked Example: Fairness Metrics in Recruitment

Suppose a recruitment screening system is evaluated for two groups. Group A has 100 applicants, of whom 40 are qualified; the system selects 30 qualified applicants and rejects 10 qualified applicants. It also selects 12 unqualified applicants and rejects 48. Group B has 100 applicants, of whom 50 are qualified; it selects 30 qualified applicants and rejects 20 qualified applicants. It selects 10 unqualified applicants and rejects 40.

For Group A:

TPR_A = 30 ÷ (30 + 10) = 0.75, or 75%.

For Group B:

TPR_B = 30 ÷ (30 + 20) = 0.60, or 60%.

The difference in true-positive rates is 15 percentage points. The system therefore does not satisfy equal opportunity under this evaluation. For false-positive rates:

FPR_A = 12 ÷ (12 + 48) = 0.20, or 20%.

FPR_B = 10 ÷ (10 + 40) = 0.20, or 20%.

The false-positive rates are equal in this sample, but that does not establish overall fairness. The organisation should investigate labels, job-relatedness, sample uncertainty, accessibility, the consequences of selection and whether the chosen metric reflects the applicable law and organisational values. It should not simply adjust thresholds to improve a metric without examining underlying causes and legal implications.

Case Study: Illustrative Public Benefits Eligibility System

Background

A fictional regional authority introduces an AI system to prioritise applications for emergency housing assistance. The system uses household income, previous applications, rental history, dependants and administrative records. Caseworkers receive a priority score and are expected to process the highest scores first.

Problem

After deployment, an advocacy organisation reports that applicants with unstable housing and limited digital access receive lower scores. The authority also discovers that some applicants were incorrectly linked to previous addresses. Caseworkers have little time to inspect the underlying data and rarely override the score.

Analysis

The system has a high-impact public-sector context. The main risks are inaccurate data, proxy discrimination, exclusion of people with incomplete records, automation bias, inadequate explanation and lack of effective appeal. The issue is not merely model accuracy: the workflow converts a prioritisation tool into a de facto eligibility gate.

Approach and solution

The authority should immediately suspend automatic rejection and ensure access to human caseworkers. It should correct identity matching and permit applicants to submit alternative evidence. A multidisciplinary review should examine the objective, data provenance, subgroup outcomes, disability and language accessibility, and the effect of digital exclusion. The authority should publish an understandable description of the system and provide reasons for individual outcomes.

The model may be redesigned as a triage aid rather than a dispositive decision-maker. Caseworkers should receive training on automation bias, have authority to override scores and record reasons. Monitoring should include error rates, processing times, successful appeals and outcomes for groups likely to experience incomplete administrative records. Procurement contracts should provide audit access and require notification of model changes.

Outcome

In this illustrative case, the authority restores a manual route, improves data correction, narrows the system’s permitted purpose and establishes periodic independent review. The result is not necessarily a more complex model; it is a more accountable decision process.

Learning points

1. A lawful objective does not legitimise an unsuitable implementation.
2. Administrative data may be incomplete in systematically unequal ways.
3. Human oversight requires time, authority and usable reasons.
4. Appeals and correction are substantive safeguards, not public-relations features.
5. Impact must be evaluated at workflow and institutional levels.

Challenges and Limitations

AI governance faces several persistent difficulties. First, fairness metrics may conflict and may conceal low sample sizes or measurement error. Second, explainability techniques can provide persuasive but non-causal explanations. Third, audit evidence may become obsolete when a model, prompt, data source or user population changes. Fourth, small organisations may lack specialist expertise and bargaining power over vendors. Fifth, regulatory obligations differ across jurisdictions and sectors, creating complexity for international deployments.

There are also limits to technical fixes. Bias cannot always be removed through better data; some harms arise from the decision itself. Privacy cannot be reduced to anonymisation. Human review cannot compensate for impossible workloads. Transparency without remedy may merely expose unfairness without correcting it. Responsible innovation therefore requires institutional capacity, affected-person participation and willingness not to deploy a system when risks cannot be controlled.

Professional and Managerial Implications

IT professionals should treat model cards, data sheets, test reports, audit logs, change records and incident reports as engineering and governance artefacts. Managers should allocate ownership, budget for monitoring, define risk appetite and ensure that commercial pressure does not bypass approval. Legal and compliance professionals should translate broad principles into operational requirements and keep a current register of applicable law. Procurement teams should avoid vague promises such as “bias-free AI” and demand verifiable evidence, audit rights and meaningful remedies.

Boards and senior executives should ask: What decision is being automated? Who may be harmed? What evidence supports reliability? Which groups were tested? What happens when the system is wrong? Who can stop it? How can an affected person obtain correction and remedy? These questions connect innovation with institutional responsibility.

Contemporary Developments and Best Practices

Current developments include foundation models, multimodal systems, autonomous agents, synthetic data, watermarking and algorithmic impact assessments. These technologies increase capability but also make boundaries of responsibility less clear. Organisations should maintain an AI policy that distinguishes experimentation from production use, define approved and prohibited applications, and require heightened review for systems affecting rights, safety, employment, education, health, credit or access to public services.

A concise best-practice checklist is:

• establish an AI inventory and accountable owner;
• define purpose, legal basis and prohibited uses;
• conduct proportionate impact assessments;
• document data provenance, limitations and licences;
• test accuracy, robustness, accessibility and subgroup outcomes;
• implement security, privacy and least-privilege controls;
• provide meaningful human oversight and appeal;
• contractually govern vendors and model changes;
• monitor drift, incidents, complaints and real-world outcomes;
• retain evidence sufficient for audit and remedy; and
• review or retire systems when risks exceed benefits.

Chapter Summary

AI and automated decision systems are sociotechnical arrangements whose risks arise from data, models, workflows, institutions and deployment contexts. Responsible innovation requires accountability, legality, necessity, proportionality, fairness, transparency, explainability, privacy, security, resilience and meaningful human oversight.

Bias can enter during problem formulation, data collection, labelling, modelling, deployment and feedback. Fairness metrics such as true-positive and false-positive rates are useful diagnostic tools but do not provide a universal definition of justice. Privacy and intellectual-property analysis must address prompts, training data, outputs, logs, licences and contractual rights. Generative AI adds hallucination, prompt-injection, data-leakage, provenance and agentic-action risks.

Risk-based legislation, including the EU AI Act, should be considered alongside data-protection law, sectoral duties and voluntary frameworks such as NIST AI RMF, ISO/IEC 42001 and the OECD AI Principles. Effective governance operates throughout the life cycle: inventory, assess, design, test, approve, monitor, remedy and retire. The central professional obligation is not merely to build accurate systems, but to build decision processes that are lawful, contestable, secure and worthy of trust.

Key Terms

Algorithmic accountability: The ability to identify responsibility, provide evidence, review decisions and deliver remedies for algorithmic outcomes.

Automated decision system: A system that uses computational procedures to support or make decisions affecting people or organisations.

Automation bias: The tendency to over-rely on automated outputs, even when they are inaccurate.

Bias: A systematic pattern that produces inaccurate, unfair or disadvantageous outcomes.

Calibration: The extent to which predicted probabilities correspond to observed frequencies.

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

Demographic parity: A fairness criterion comparing positive-outcome rates across groups.

Explainability: The degree to which system behaviour or outputs can be understood by an appropriate audience.

Fairness: A contextual normative assessment of whether outcomes, processes and treatment are just and non-discriminatory.

Generative AI: AI capable of producing new content such as text, images, audio, video or code.

Human oversight: Competent and empowered human involvement capable of detecting, preventing or correcting harmful outcomes.

Model drift: Degradation or change in model performance as data, behaviour or environments change.

Prompt injection: An attack in which instructions in input or retrieved content manipulate a generative system.

Proportionality: The principle that an intervention should be suitable, necessary and balanced against its effects.

Sociotechnical system: A system formed by technology, people, organisations, processes and social context.

Transparency: Disclosure about a system’s purpose, operation, data, limitations and governance.

True-positive rate: The proportion of actual positive cases correctly identified, calculated as TP ÷ (TP + FN).

Vendor governance: Contractual and operational controls governing third-party AI providers and suppliers.

Review Questions

5 Short Answer Questions

1. What distinguishes an automated decision system from a conventional information system?
2. Define algorithmic accountability and identify two accountability records.
3. Why does removing a protected attribute not necessarily remove discrimination?
4. What is automation bias, and how can interface design reduce it?
5. Identify three risks specific to generative-AI deployments.

5 Long Answer Questions

1. Explain why AI governance must address the model, the complete workflow and the institution.
2. Compare transparency and explainability, including their different audiences and limitations.
3. Discuss how privacy, intellectual property and contractual governance interact in enterprise generative-AI use.
4. Explain the strengths and limitations of fairness metrics such as demographic parity, equal opportunity and equalised odds.
5. Describe a complete life-cycle governance programme for a high-impact AI system.

3 Analytical / Application Questions

1. A bank uses an AI model to rank applicants for manual credit review. The model has equal overall accuracy across two groups but materially higher false-negative rates for one group. Analyse the governance issues and recommend tests, controls and remedies.
2. A university deploys a generative chatbot connected to student records and fee-payment tools. Design a risk assessment covering privacy, security, human oversight, hallucination and agentic action.
3. An organisation purchases a vendor system that claims to be “bias-free” but refuses to disclose training data or permit independent testing. Evaluate whether deployment should proceed and formulate contractual and governance requirements.

Multiple Choice Questions with Answers

Question 1

Which statement best describes algorithmic accountability?

A. The belief that algorithms are responsible for their own outcomes

B. The ability to identify responsibility, review evidence and provide remedies

C. The requirement that every AI model be open source

D. The use of accuracy as the only governance measure

Correct Answer: B. The ability to identify responsibility, review evidence and provide remedies

Question 2

For a binary classifier, the true-positive rate is calculated as:

A. TP ÷ (TP + FN)

B. FP ÷ (FP + TN)

C. TP ÷ (TP + FP)

D. TN ÷ (TN + FN)

Correct Answer: A. TP ÷ (TP + FN)

Question 3

Why can removing a protected attribute fail to eliminate discrimination?

A. Models cannot operate without protected attributes

B. The attribute may be replaced by random noise only

C. Other variables may act as correlated proxies

D. Protected attributes always improve accuracy

Correct Answer: C. Other variables may act as correlated proxies

Question 4

Which condition is necessary for meaningful human oversight?

A. The reviewer must approve every output automatically

B. The reviewer must have competence and authority to intervene

C. The system must hide uncertainty from the reviewer

D. The reviewer must be responsible for model training

Correct Answer: B. The reviewer must have competence and authority to intervene

Question 5

Which is a generative-AI-specific risk?

A. Hallucinated content presented as factual

B. A database having a primary key

C. A network using encryption

D. A printer losing power

Correct Answer: A. Hallucinated content presented as factual

Question 6

What is the main purpose of data provenance?

A. To increase model complexity

B. To record the origin, transformation and permitted use of data

C. To guarantee that all outputs are copyrighted

D. To replace access controls

Correct Answer: B. To record the origin, transformation and permitted use of data

Question 7

Which statement about fairness metrics is most accurate?

A. One metric universally proves that a system is fair

B. Fairness metrics are irrelevant to legal analysis

C. Different metrics can conflict and must be interpreted contextually

D. Equal overall accuracy guarantees equal treatment

Correct Answer: C. Different metrics can conflict and must be interpreted contextually

Question 8

Which framework is organised around Govern, Map, Measure and Manage?

A. NIST AI Risk Management Framework

B. A copyright registration system

C. A database normalisation standard

D. A programming language specification

Correct Answer: A. NIST AI Risk Management Framework

Question 9

What is prompt injection?

A. Compressing a prompt to reduce storage costs

B. Manipulating a generative system through embedded instructions

C. Encrypting a prompt before transmission

D. Removing all instructions from a model

Correct Answer: B. Manipulating a generative system through embedded instructions

Question 10

Which is the strongest reason to treat AI governance as a life-cycle process?

A. Model and deployment risks can change after initial approval

B. AI systems never require monitoring after release

C. Testing is useful only before data collection

D. Retirement eliminates the need for documentation

Correct Answer: A. Model and deployment risks can change after initial approval