Responsible AI, Governance, and Future Readiness

Introduction

Artificial intelligence creates entrepreneurial opportunities in automation, prediction, personalisation, scientific discovery and new venture formation. It also creates risks that can damage individuals, communities, organisations and the venture itself. A biased recruitment tool may exclude qualified applicants; an insecure model may expose confidential data; an inaccurate medical recommendation may cause physical harm; and an opaque automated decision may be impossible for a customer to challenge.

Responsible AI is therefore not merely an ethics exercise. It is a managerial, technical, legal and strategic capability. Entrepreneurs must design systems that create value while respecting rights, controlling risks and remaining adaptable as technologies and regulations change. Responsible practice should begin before a model is trained and continue through deployment, monitoring, retirement and post-incident learning.

This chapter develops an integrated approach to fairness, transparency, explainability, accountability, privacy, security, safety, sustainability, intellectual property, audit controls and human oversight. It then applies foresight and scenario planning to future readiness.

Learning Objectives

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

1. Distinguish the principal dimensions of responsible AI and explain their interaction.
2. Identify ethical, legal, operational and strategic risks in an AI-enabled venture.
3. Design proportionate governance, audit and human-oversight controls.
4. Evaluate fairness, privacy, security, safety, sustainability and intellectual-property concerns.
5. Apply a risk-based approach to regulation and venture decision-making.
6. Use foresight and scenario planning to prepare a venture for uncertain technological and regulatory futures.

1. Responsible AI as a Venture Capability

Responsible AI means developing, deploying and governing AI systems so that they are beneficial, lawful, contestable, secure, safe and socially legitimate. It is broader than avoiding obviously harmful uses. It includes asking whether a proposed use is appropriate, whether affected people understand it, whether the system performs adequately across groups, and whether the organisation can detect and correct failures.

A useful distinction is between principles, controls and evidence:

• Principles state what the venture values, such as fairness, privacy and safety.
• Controls translate values into procedures, technical safeguards and decision rights.
• Evidence demonstrates that controls operated effectively, through records, tests, audits and incident reports.

Principles without controls become aspirational statements. Controls without evidence cannot be trusted. Evidence without clear principles can measure activity without measuring what matters.

Responsible AI also involves trade-offs. Increasing privacy may reduce access to data. A highly accurate model may be less interpretable. More human review may improve accountability but increase cost and delay. Entrepreneurs should not conceal these trade-offs; they should document them, justify decisions and select controls proportionate to potential harm.

Important Terminology

AI system: A machine-based system that generates predictions, recommendations, content or decisions from inputs.

High-impact use: A use that can materially affect rights, opportunities, safety, health, finances or access to essential services.

Model risk: The possibility that a model is unsuitable, inaccurate, unstable or misused.

Data provenance: Evidence about where data came from, how it was collected, transformed and used.

Contestability: The ability of an affected person to question, appeal or seek review of an AI-assisted outcome.

Human-in-the-loop: A human actively reviews or approves a system output before a consequential action.

Human-on-the-loop: A human supervises an operating system and intervenes when predefined conditions arise.

Drift: Change over time in data, relationships, user behaviour or model performance.

Red teaming: Structured adversarial testing intended to expose weaknesses, misuse or harmful outputs.

2. Fairness and Non-Discrimination

Fairness concerns whether an AI system produces unjustified disadvantages or benefits for individuals or groups. Discrimination may arise from biased historical data, under-representation, proxy variables, inappropriate labels, unequal measurement quality or the context in which an otherwise accurate model is used.

Fairness is not one universal statistical property. Common perspectives include:

• Individual fairness: Similar individuals should receive similar treatment where relevant similarities are defined appropriately.
• Group fairness: Performance or outcomes should not systematically disadvantage protected or vulnerable groups.
• Procedural fairness: People should receive notice, consistent procedures, reasons and an opportunity to challenge decisions.
• Substantive fairness: The social consequences of a system should be examined, not merely its numerical parity.

Different fairness criteria can conflict, particularly when groups have different base rates or when error costs differ. Consequently, an entrepreneur should not select a metric mechanically. The venture must first identify the decision, the affected groups, the harms and the legitimate business objective.

A practical fairness assessment asks:

1. Which people may be affected directly or indirectly?
2. Are protected characteristics or sensitive attributes present, and are they needed for auditing?
3. Does the dataset represent the intended population and context?
4. Are error rates, access rates and override rates materially different across groups?
5. Could a feature act as a proxy for a protected characteristic?
6. What remedy is available when the system is wrong?

Fairness controls include representative sampling, data-quality checks, subgroup testing, threshold review, accessibility design, human appeals and monitoring after deployment. Removing a protected attribute does not guarantee fairness because other variables may encode it indirectly.

3. Transparency, Explainability and Contestability

Transparency concerns whether relevant information about a system is disclosed. It may include the system’s purpose, data sources, provider, limitations, operating context and responsible decision-maker. Explainability concerns whether a system’s operation or output can be meaningfully explained to relevant audiences.

Explanations should be audience-specific. A customer may need a plain-language reason and appeal route. A regulator may require records of data, controls and risk assessment. A technical reviewer may need model architecture, validation results and feature behaviour. A founder needs enough information to decide whether the system is fit for purpose.

There are two broad forms of explanation:

• Global explanation: How the system generally behaves across cases.
• Local explanation: Why a particular output was produced for a particular case.

A feature-importance ranking is not necessarily a causal explanation. Correlation does not prove that changing a feature would change the outcome. Explanations can also be misleading if they are simplified, unstable or generated after the fact. Entrepreneurs should test whether users can understand and act on explanations rather than treating interpretability as a decorative interface feature.

Minimum transparency documentation commonly includes intended use, prohibited use, data lineage, evaluation population, performance limitations, known failure modes, version history, responsible owners and contact or appeal channels.

4. Accountability and Governance Architecture

Accountability means that identifiable people and organisational bodies are answerable for decisions and outcomes. An AI system is not an accountable actor: responsibility remains with the venture, its leaders and relevant suppliers.

A small venture can establish a proportionate governance architecture without creating a large bureaucracy. Key roles may include:

• Board or founders: Set risk appetite, approve high-impact uses and allocate resources.
• Product owner: Defines purpose, users, success criteria and unacceptable use.
• Technical owner: Manages data, model development, validation, monitoring and change control.
• Risk or compliance lead: Maintains the risk register and checks legal and policy obligations.
• Domain expert: Assesses whether outputs make sense in the relevant professional context.
• Human reviewer: Exercises judgement, records decisions and handles escalation.
• Supplier manager: Verifies external model providers and contractual responsibilities.

A responsibility matrix should identify who is responsible for performing a task, accountable for its outcome, consulted before decisions and informed afterward. Governance should cover the complete lifecycle: proposal, design, data preparation, testing, approval, deployment, monitoring, incident response and retirement.

A model card or system record should state intended purpose, users, training-data characteristics, evaluation results, limitations, risks, version and approval status. A change to training data, model version, prompt, retrieval source or user population may require renewed assessment.

5. Privacy and Data Governance

Privacy is concerned with appropriate collection, use, disclosure, retention and deletion of information about identifiable or potentially identifiable people. Data protection obligations vary by jurisdiction, but sound practice generally requires a lawful and transparent purpose, data minimisation, accuracy, security, retention limits and respect for individual rights.

Entrepreneurs should distinguish:

• Personal data: Information relating to an identified or identifiable person.
• Sensitive or special-category data: Data receiving stronger protection in many legal systems, such as health, biometric or certain demographic information.
• Anonymised data: Data from which identification is not reasonably possible under the relevant standard.
• Pseudonymised data: Data where identifiers are replaced but re-identification remains possible with additional information.

Removing names does not automatically anonymise a dataset. Several harmless-looking variables can be combined to re-identify people. Privacy risk also arises when a model memorises or reveals training information.

A privacy-by-design process should define purpose and necessity before collecting data; document provenance and permissions; restrict access; encrypt data in transit and at rest; separate production and development environments; establish retention and deletion schedules; and assess vendors. Privacy-enhancing techniques such as aggregation, access controls, de-identification, federated learning or differential privacy may reduce risk, but each has limitations and must be validated in context.

6. Security, Safety and Resilience

Security protects systems, data and users from unauthorised access, manipulation and disruption. Safety concerns whether the system causes unacceptable physical, psychological, financial or social harm during intended or reasonably foreseeable use. A system can be secure but unsafe, or safe in normal operation but vulnerable to attack.

AI-specific security threats include prompt injection, data poisoning, model theft, adversarial inputs, insecure plugins, supply-chain compromise, unauthorised fine-tuning and leakage of confidential information. General cybersecurity controls remain essential: least-privilege access, strong authentication, network segmentation, secure development, dependency management, logging, backup and incident response.

Safety engineering requires hazard identification and controls. Entrepreneurs should ask:

1. What could go wrong?
2. Who could be harmed and how severely?
3. How likely is the harm under normal, unusual and adversarial conditions?
4. Can the system detect uncertainty or out-of-distribution inputs?
5. What safe fallback or shutdown mechanism exists?
6. Who is authorised to intervene?

For high-consequence applications, useful safeguards include bounded operating conditions, conservative defaults, independent validation, fail-safe states, rate limits, staged rollout, simulation, stress testing and mandatory human escalation. A disclaimer is not a substitute for engineering controls.

7. Sustainability and Resource Responsibility

AI systems consume energy, water, hardware and human labour. Sustainability should be evaluated across the lifecycle, including data collection, training, inference, hardware manufacture, transportation and disposal. Entrepreneurs should consider whether AI is necessary for the task, whether a smaller model is sufficient, and whether computation can be reduced through caching, batching, efficient architectures or selective invocation.

A basic operational estimate of energy use is:

where E_total is energy in watt-hours when P is average power in watts and t is operating time in hours. The formula assumes power is reasonably constant. If power varies over time, the more general relationship is:

where P(t) is power at time t. This estimate excludes embodied energy unless that is separately measured. A venture should report the boundary, assumptions and data quality rather than presenting an estimate as an exact fact.

Sustainability also includes social effects: the quality and compensation of data-labelling work, accessibility, effects on employment and the distribution of benefits. Efficiency should not be pursued by reducing safety testing or transferring hidden costs to vulnerable communities.

8. Regulation and Intellectual Property

AI regulation is jurisdiction-specific and rapidly developing. Ventures must distinguish legal requirements from voluntary standards and internal policies. Depending on the use, relevant obligations may concern data protection, consumer protection, equality and anti-discrimination, product safety, employment, financial services, health, cybersecurity, copyright and sector licensing.

A risk-based regulatory approach classifies uses according to potential impact. Prohibited or unacceptable practices may be restricted; high-risk systems may require stronger documentation, testing, human oversight, monitoring and incident reporting; lower-risk uses may require transparency and ordinary controls. Classification should be based on actual use and effect, not only on the marketing label attached to the product.

Entrepreneurs should maintain a regulatory register containing jurisdictions, applicable rules, responsible owners, required evidence, review dates and escalation procedures. Legal review is especially important before processing sensitive data, making consequential decisions, exporting technology or serving regulated sectors.

Intellectual-property issues include:

• ownership and licensing of training, retrieval and evaluation data;
• copyright and database rights in inputs and outputs;
• trade secrets in prompts, workflows, datasets and model weights;
• patents or freedom-to-operate concerns for technical inventions;
• contractual restrictions imposed by model and data suppliers;
• rights of publicity, trademarks and confidentiality.

Do not assume that an AI-generated output is automatically protectable or that a publicly available dataset is free to use. Keep provenance records, review licences, obtain permissions where necessary and avoid placing confidential client information into external systems without contractual and technical safeguards. Human contribution, originality and jurisdiction can affect intellectual-property outcomes.

9. Audit Controls and Human Oversight

An audit trail allows an organisation to reconstruct what happened, who approved it, which model and data version were used, what output was produced and what action followed. Appropriate records may include access logs, model versions, prompts or input categories, validation results, overrides, complaints, incidents, vendor attestations and corrective actions. Logging must itself respect privacy and security requirements.

A practical control framework has four layers:

Human oversight is meaningful only when the reviewer has competence, time, authority, relevant information and a genuine ability to reject or correct the output. A human who merely clicks approval under workload pressure provides little accountability. Oversight should define escalation thresholds, review sampling, override recording, conflict-of-interest controls and emergency shutdown authority.

Monitoring should examine technical and social indicators, such as accuracy, subgroup error rates, false positives, complaints, override frequency, drift, latency, security alerts and energy use. Metrics should trigger action rather than become passive reporting. Possible actions include investigation, threshold change, retraining, restricted use, rollback or withdrawal.

10. A Practical Responsible-AI Lifecycle

A venture can operationalise responsible AI through the following lifecycle:

1. Justify the use: Define the problem, affected stakeholders, alternatives and expected benefit.
2. Classify risk: Consider severity, scale, reversibility, vulnerability and regulatory exposure.
3. Map data and dependencies: Record sources, permissions, vendors, transformations and retention.
4. Design safeguards: Specify fairness, privacy, security, safety, accessibility and human-review requirements.
5. Build and test: Use representative evaluation data, subgroup tests, adversarial tests and domain review.
6. Approve deployment: Require documented acceptance criteria, owners, limitations and rollback conditions.
7. Monitor in operation: Track performance, drift, incidents, complaints and unintended effects.
8. Respond and learn: Investigate incidents, notify as required, remedy affected people and update controls.
9. Retire responsibly: Remove access, preserve required records, delete data where appropriate and communicate the change.

A simple risk-prioritisation tool can use severity and likelihood categories. It should not be treated as a scientifically precise probability model. A low-likelihood event with catastrophic consequences may require stronger controls than a frequent but minor inconvenience.

11. Foresight and Scenario Planning

Future readiness is the capacity to remain viable under plausible changes in technology, regulation, markets, social expectations and infrastructure. Forecasting attempts to predict what will happen; foresight explores what could happen and what the venture should do under each possibility.

Useful foresight methods include horizon scanning, trend analysis, expert interviews, technology road mapping, weak-signal monitoring, backcasting and scenario planning. Scenario planning is especially useful when uncertainty is high and decisions are difficult to reverse.

A practical process is:

1. Define the focal question, such as “How should the venture deliver trustworthy AI services over the next five years?”
2. Identify driving forces: regulation, model capability, compute cost, customer trust, labour markets, cyber threats and environmental constraints.
3. Separate predetermined trends from critical uncertainties.
4. Select two important uncertainties and construct a 2×2 scenario matrix.
5. Describe each scenario’s customers, competitors, costs, risks and opportunities.
6. Test the venture’s current strategy against every scenario.
7. Identify robust actions, contingent actions and early-warning indicators.
8. Assign owners and review the scenarios periodically.

For example, one axis may represent strict versus permissive regulation, while another represents concentrated versus widely accessible advanced models. The resulting scenarios might include regulated concentration, open experimentation, fragmented compliance and commoditised capability. The objective is not to choose the most attractive story. It is to identify strategies that remain sensible across several futures.

Robust actions are valuable in nearly every scenario, such as maintaining data provenance, modular system architecture, security controls and staff competence. Contingent actions are activated by signals, such as obtaining a sector certification when a draft rule becomes likely. No-regret actions improve resilience regardless of the future, such as documenting model limitations and creating an incident-response plan.

12. Illustrative Case: CareRoute Health Services

Background

CareRoute is a fictional early-stage venture offering an AI tool that prioritises follow-up appointments for patients in community clinics. The system uses clinical and administrative information to identify patients who may need earlier contact.

Problem

An initial pilot showed good overall predictive performance, but nurses observed that the system generated fewer priority recommendations for patients who had irregular appointment histories. Many of those patients had language, transport or access barriers. The venture also planned to use a third-party generative AI service to summarise records, creating confidentiality and supplier-governance concerns.

Analysis

The product involved a high-impact health use. The fairness issue was not necessarily caused by an explicit protected attribute; appointment history acted as a proxy for access and socioeconomic conditions. A high overall accuracy figure concealed unequal errors. Privacy risks included excessive data transfer, unclear retention and possible disclosure through prompts. Safety risks arose if staff treated the priority score as a diagnosis rather than a triage aid.

Approach and Solution

CareRoute paused expansion and established a governance review. It:

• defined the tool as decision support, not an autonomous clinical decision-maker;
• conducted subgroup and error analysis with clinicians and patient representatives;
• added an access-barrier indicator and tested whether it improved recall without creating new harms;
• created a human-review rule for low-confidence and high-risk cases;
• contracted a provider that prohibited secondary use of clinic data and specified deletion and breach duties;
• separated identifiable records from development data and limited access by role;
• introduced audit logs, model-version records and a rollback procedure;
• informed patients that an AI-assisted prioritisation tool was used and provided a correction route;
• monitored override rates, complaints, subgroup performance and adverse incidents.

Outcome

The revised pilot did not maximise a single accuracy metric. Instead, it produced a more defensible service: clinicians understood its limitations, patients had a route to challenge outcomes, and the venture could demonstrate its controls to clinic partners. The system remained subject to periodic review because changes in patient populations, workflows or regulation could invalidate earlier assumptions.

Learning Points

The case illustrates that responsible AI is socio-technical. Better model performance alone would not solve the problem. Fairness required contextual knowledge; privacy required contractual and technical controls; safety required role boundaries and escalation; accountability required named owners and records. The venture also protected its commercial prospects because trustworthy governance became part of its value proposition.

13. Challenges and Limitations

Responsible AI frameworks have important limitations. Fairness metrics can conflict and may encourage optimisation of what is measurable rather than what is just. Explanations may create false confidence. Audits are snapshots and cannot guarantee future performance. Voluntary principles may be weaker than enforceable rights. Small ventures may lack specialist staff, while external vendors may withhold information about proprietary models. Human review can reproduce bias or become ineffective under time pressure. Sustainability estimates may be incomplete because suppliers do not disclose energy, water or embodied impacts.

These limitations do not justify inaction. They require humility, documentation, stakeholder participation, independent challenge and continuous monitoring. The appropriate question is not whether a system is perfectly responsible; it is whether the venture has recognised material risks, reduced them proportionately, provided remedies and learned from evidence.

14. Best Practices for Entrepreneurs

1. Begin with the use case and potential harm, not with an attractive model.
2. Maintain a living AI-system inventory and risk register.
3. Assign named accountability for every consequential system.
4. Keep data, model, prompt, supplier and decision provenance.
5. Test performance across relevant groups and real operating conditions.
6. Make limitations and appeal routes understandable to affected users.
7. Use least privilege, secure development and documented incident response.
8. Require meaningful human oversight for high-impact decisions.
9. Include audit, deletion, security, liability and change-notification clauses in supplier contracts.
10. Monitor emerging regulation, standards, threats, customer expectations and environmental constraints.
11. Prefer modular, reversible deployments and staged pilots.
12. Treat governance evidence as an entrepreneurial asset that supports trust, investment and market access.

Chapter Summary

Responsible AI is a lifecycle capability combining ethical principles, technical controls, organisational accountability and verifiable evidence. Fairness requires attention to context, representation, error distribution and contestability rather than reliance on one metric. Transparency and explainability must serve the needs of customers, reviewers, regulators and technical teams. Privacy requires purpose limitation, minimisation, security, retention control and careful treatment of re-identification risk. Security protects systems from attack, while safety manages foreseeable harm through testing, bounded operation, escalation and fail-safe design.

Governance should assign clear roles, preserve audit trails and provide meaningful human oversight. Regulation and intellectual property are jurisdiction- and context-dependent, requiring continuous review, provenance and competent legal advice. Sustainability includes energy, water, hardware, labour and social impacts. Finally, foresight and scenario planning help ventures prepare for uncertain changes in capability, regulation, competition and public expectations. Future-ready ventures pursue innovation and responsibility together.

Key Terms

Responsible AI: Development and use of AI that is lawful, fair, safe, secure, explainable and accountable.

Fairness: The avoidance of unjustified disadvantage and the provision of appropriate procedural treatment.

Transparency: Disclosure of relevant information about an AI system and its operation.

Explainability: The ability to provide understandable reasons for system behaviour or outputs.

Accountability: The assignment of responsibility and answerability for decisions and outcomes.

Contestability: The ability to question, appeal or seek review of an AI-assisted decision.

Data provenance: A record of data origin, processing, permissions and use.

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

Pseudonymisation: Replacing identifiers while retaining a possibility of re-identification.

Model risk: The possibility that a model is inaccurate, unsuitable, unstable or misused.

Data drift: Change in the distribution of inputs over time.

Concept drift: Change in the relationship between inputs and the outcome being predicted.

Red teaming: Adversarial testing to identify weaknesses, misuse and harmful outputs.

Human-in-the-loop: Human review or approval before a consequential action.

Audit trail: Records that allow decisions, versions, actions and responsible parties to be reconstructed.

Safety case: Structured evidence that a system is acceptably safe for a defined use.

Risk-based regulation: Regulation whose obligations increase with potential impact and harm.

Intellectual property: Legal rights relating to creations, inventions, confidential information and branding.

Scenario planning: Exploration of plausible alternative futures to improve strategic resilience.

Horizon scanning: Systematic monitoring of emerging trends, threats and opportunities.

Review Questions

5 Short Answer Questions

1. Why are principles, controls and evidence all necessary for responsible AI?
2. Distinguish transparency from explainability.
3. Give two reasons why removing a protected attribute may not eliminate discrimination.
4. What makes human oversight meaningful rather than symbolic?
5. Distinguish data drift from concept drift.

5 Long Answer Questions

1. Explain how fairness, privacy and safety can conflict in an entrepreneurial AI application. Propose a method for managing the trade-offs.
2. Design a governance structure for a small venture deploying an AI system in a high-impact domain.
3. Discuss the role of audit trails, model records and supplier contracts in establishing accountability.
4. Explain the major intellectual-property risks associated with training and deploying generative AI systems.
5. Evaluate how sustainability should be incorporated into AI product and architecture decisions.

3 Analytical / Application Questions

1. A lending startup reports equal overall accuracy across demographic groups but receives complaints that one group is frequently rejected incorrectly. Analyse what additional evidence and controls are required before deployment.
2. Construct a scenario-planning approach for a venture whose business depends on an external foundation-model provider. Identify two critical uncertainties, four scenarios and three robust actions.
3. Design an incident-response and human-oversight procedure for an AI system that recommends urgent healthcare follow-up but occasionally produces low-confidence outputs.

Multiple Choice Questions with Answers

Question 1

Which statement best describes responsible AI as presented in the chapter?

A. Maximising model accuracy regardless of social consequences

B. Using AI only when no human is involved

C. Combining beneficial, lawful, fair, secure, safe and accountable AI practices

D. Publishing all source code and training data publicly

Correct Answer: C. Combining beneficial, lawful, fair, secure, safe and accountable AI practices

Question 2

Why does removing a protected attribute not guarantee fairness?

A. Models cannot process numerical variables

B. Other variables may act as proxies for the protected attribute

C. Protected attributes are always required for prediction

D. Accuracy automatically becomes zero

Correct Answer: B. Other variables may act as proxies for the protected attribute

Question 3

What is the main distinction between transparency and explainability?

A. Transparency concerns relevant system information, while explainability concerns understandable reasons for outputs

B. Transparency applies only to hardware, while explainability applies only to software

C. Transparency guarantees fairness, while explainability guarantees privacy

D. They are identical terms with no practical distinction

Correct Answer: A. Transparency concerns relevant system information, while explainability concerns understandable reasons for outputs

Question 4

Which condition is essential for meaningful human oversight?

A. The reviewer must approve every output instantly

B. The reviewer must have authority, competence and a genuine ability to intervene

C. The reviewer should not see the model’s limitations

D. The reviewer should be responsible only after an incident

Correct Answer: B. The reviewer must have authority, competence and a genuine ability to intervene

Question 5

If a system operates at a constant average power of 200 watts for 5 hours, what is its approximate energy use?

A. 40 Wh

B. 205 Wh

C. 1,000 Wh

D. 10,000 Wh

Correct Answer: C. 1,000 Wh

Question 6

Which record most directly supports reconstruction of an AI-assisted decision?

A. A marketing slogan

B. An audit trail containing versions, approvals and relevant actions

C. A prediction made by an unrelated model

D. A general statement that the system is innovative

Correct Answer: B. An audit trail containing versions, approvals and relevant actions

Question 7

What is a principal purpose of red teaming?

A. To increase advertising reach

B. To remove the need for monitoring

C. To expose weaknesses, misuse and harmful outputs through adversarial testing

D. To guarantee that regulation will not change

Correct Answer: C. To expose weaknesses, misuse and harmful outputs through adversarial testing

Question 8

Which statement about pseudonymised data is correct?

A. It can never be re-identified

B. It replaces identifiers but re-identification may remain possible

C. It is always publicly available

D. It contains no personal information under any circumstances

Correct Answer: B. It replaces identifiers but re-identification may remain possible

Question 9

What is the primary aim of scenario planning?

A. To predict one certain future precisely

B. To explore plausible futures and identify resilient actions

C. To replace legal compliance with imagination

D. To eliminate all entrepreneurial risk

Correct Answer: B. To explore plausible futures and identify resilient actions

Question 10

Which is an example of a robust or no-regret action for an AI venture?

A. Ignoring model limitations until a competitor raises them

B. Maintaining data provenance and an incident-response plan

C. Deploying only the largest available model

D. Removing all human review to reduce costs

Correct Answer: B. Maintaining data provenance and an incident-response plan