Part III — Evaluating Risk, Intelligent Technologies, Digital Systems, Biotechnology and Human-Centred Technologies

Chapter 10. Evaluating Artificial Intelligence and Intelligent Technologies — Part 3 of 3

10.27 Case Study: Generative AI in Higher Education

A university introduces generative AI to support:

student tutoring;

administrative communication;

academic writing support;

curriculum development;

research assistance.

Potential benefits include accessibility, productivity and personalized learning.

Potential risks include:

inaccurate information;

academic integrity concerns;

privacy;

dependence;

inequitable access;

inappropriate automation;

loss of human interaction;

intellectual-property questions.

Evaluation should therefore involve students, academics, administrators, information-security specialists, legal personnel and institutional leadership.

The higher-education case demonstrates particularly clearly that generative AI cannot be evaluated through a single institutional performance measure. Universities are complex socio-technical environments in which technology interacts with teaching, learning, research, assessment, administration, governance, academic freedom, professional judgment, student development, and institutional reputation.

The first question should concern institutional purpose. The university should determine why generative AI is being introduced and what educational or organizational problem it is intended to address. The fact that a technology is capable of generating text, explanations, summaries, questions, code, images, or other materials does not automatically establish that those capabilities improve educational outcomes.

For student tutoring, the university may intend to provide additional learning support outside conventional teaching hours. The evaluator should therefore examine whether the system provides accurate and pedagogically appropriate assistance, whether it encourages student learning rather than simply producing answers, and whether students understand when the system may be unreliable.

Personalized learning presents another evaluation question. Generative AI may adapt explanations to individual students, provide examples at different levels of difficulty, assist with language barriers, and generate additional practice material. However, personalization should be evaluated against actual learning outcomes rather than simply against the amount of content produced.

The evaluator should therefore distinguish between activity and educational value. A system may generate thousands of explanations or practice questions without demonstrating that students have learned more effectively.

Administrative communication provides another relatively different application. A university may use generative AI to draft emails, summarize policies, answer frequently asked questions, prepare announcements, or support student-service operations. The primary evaluation concerns may then include factual accuracy, consistency, privacy, tone, accessibility, escalation, and the need for human approval.

Where communications concern examinations, disciplinary procedures, financial matters, admissions, graduation, academic appeals, or other high-consequence issues, stronger human review may be required.

Academic writing support requires particular care. Generative AI may assist students with grammar, structure, brainstorming, organization, translation, or explanation. However, institutions need to distinguish legitimate learning support from the inappropriate outsourcing of assessed intellectual work.

The issue is therefore not simply whether students use AI, but how AI use relates to the learning objectives and assessment requirements of the programme.

Academic integrity should consequently be evaluated in relation to the purpose of assessment. If an assignment is intended to evaluate a student's ability to independently develop an argument, extensive undisclosed AI generation may undermine the validity of the assessment. If the learning objective explicitly includes responsible use of AI, however, AI-assisted work may become part of the learning process.

The university should therefore establish clear policies concerning permitted, restricted, and prohibited uses rather than relying exclusively on detection technologies.

AI detection systems themselves should be evaluated critically. A system claiming to identify AI-generated writing may produce false positives or false negatives. Consequently, such systems should not automatically be treated as definitive evidence of academic misconduct without appropriate human review and additional evidence.

Research assistance creates another major application. Generative AI can help researchers brainstorm research questions, summarize literature, organize ideas, develop coding approaches, explain statistical concepts, draft text, or support data-analysis workflows.

However, researchers must verify factual claims, references, quotations, methodological recommendations, statistical interpretations, and other outputs. Fabricated citations are particularly problematic because they can introduce apparently authoritative but nonexistent sources into academic work.

Universities should therefore promote verification practices rather than assuming that AI-generated information is reliable.

Privacy is another major institutional concern. Students and staff may enter personal, confidential, unpublished, or sensitive research information into external AI systems. Universities should therefore establish rules concerning what information may be submitted to external systems and under what conditions.

Research data require additional protection. Confidential interview transcripts, unpublished findings, proprietary datasets, identifiable participant information, intellectual property, and sensitive institutional records should not be transferred to AI services without appropriate authorization and safeguards.

Intellectual property is similarly complex. Universities should examine ownership and permitted use of AI-assisted outputs, the use of copyrighted materials, staff and student-created materials, research outputs, institutional content, and externally generated materials.

Accessibility can represent both a potential benefit and a potential risk. Generative AI may support students with language differences, learning needs, or accessibility requirements by providing alternative explanations, summaries, translations, and conversational support. However, the technology should not be assumed to work equally well across languages, disability contexts, cultural settings, or educational levels.

Equitable access should therefore be examined. Students may differ in access to devices, reliable internet connections, paid AI services, digital literacy, language proficiency, and technical support. If educational activities increasingly assume access to advanced AI tools, students with fewer resources may experience disadvantage.

The university should therefore consider whether institutional access should be provided where AI use is part of an approved learning activity and whether alternative arrangements are necessary.

Dependence is another important issue. Students may become increasingly reliant on generative AI for writing, problem-solving, coding, summarization, or idea development. The evaluation should therefore examine whether AI use strengthens or weakens students' development of independent capabilities.

This does not require assuming that all AI use creates dependence. Instead, the evaluator should examine how the technology is integrated into pedagogy. AI can potentially support learning when it encourages questioning, comparison, reflection, feedback, and revision. The same technology may contribute less effectively to learning when it simply replaces the intellectual activity students are expected to develop.

The role of academics should therefore remain important. Generative AI may reduce some routine tasks while increasing the importance of academic judgment, mentoring, assessment design, critical evaluation, and human interaction.

Curriculum development provides another application. AI can assist academics in generating learning outcomes, lesson plans, examples, discussion questions, case studies, simulations, and assessment ideas. However, generated materials require academic review because they may contain factual errors, cultural assumptions, outdated information, inappropriate examples, or pedagogical weaknesses.

Human expertise therefore remains essential in determining whether generated educational material is academically appropriate.

Student tutoring also raises the question of escalation. A generative system should be able to recognize when a student's question exceeds its intended scope or when human intervention is appropriate. For example, a student asking for general study guidance may receive automated assistance, while a student raising a complex academic appeal, safeguarding concern, serious personal issue, or specialized professional question may require referral to an appropriate human service.

This creates an important principle:

AI should not simply answer every question it receives; it should also recognize when it should not be the primary source of assistance.

The institution should therefore establish escalation pathways.

Security must also be considered. University AI systems may be connected to student-information systems, learning-management systems, library resources, research databases, email systems, or administrative platforms. If generative AI is integrated into these systems, unauthorized access or inappropriate actions could expose sensitive information.

Logging and access control should therefore be established according to the system's risk level.

The university should also consider environmental implications. Large-scale institutional use of generative AI involves computational resources, data-centre infrastructure, electricity consumption, cooling, hardware, and network resources. Environmental evaluation should therefore consider whether the educational or administrative benefits justify the resource requirements and whether lower-resource alternatives are available.

Organizational governance should establish who has authority over institutional AI use. Different responsibilities may exist for academic governance, information security, data protection, procurement, teaching and learning, research, legal compliance, and institutional strategy.

The governance structure should therefore avoid placing all responsibility on an individual academic or technology department.

The institution should also distinguish between experimentation and operational deployment. An academic may experiment with an AI tool in a low-risk environment without requiring the same governance process as an AI system that becomes integrated into student assessment, admissions, financial services, or institutional decision-making.

Pilot projects should have defined objectives, boundaries, safeguards, evidence requirements, and review points.

Evaluation should also involve multiple stakeholder groups.

Students can provide evidence concerning usability, accessibility, learning effects, trust, privacy concerns, and practical challenges.

Academics can evaluate pedagogical value, disciplinary appropriateness, academic integrity, workload, and professional implications.

Administrators can assess operational efficiency, service quality, resource requirements, and workflow integration.

Information-security specialists can examine cybersecurity, access control, data handling, system integration, and incident response.

Legal personnel can assess applicable legal and contractual requirements.

Institutional leadership can evaluate strategic alignment, resource allocation, risk appetite, accountability, and institutional consequences.

The involvement of multiple stakeholders is important because no single group possesses all the information required to evaluate the technology comprehensively.

The evaluation should also examine unintended consequences. For example, an AI system introduced to reduce administrative workload may create additional verification work. A tutoring system introduced to increase access may encourage students to bypass human academic support. A writing assistant may improve language quality while reducing opportunities for students to develop independent writing skills. An automated communication system may improve response speed while reducing the personal interaction valued by students.

These consequences illustrate the importance of distinguishing between immediate outputs and longer-term outcomes.

A useful evaluation structure can therefore be represented as:

Institutional objective → AI use case → alternatives → stakeholders → implementation → technical performance → educational outcomes → human consequences → equity → privacy → security → academic integrity → governance → monitoring → institutional learning.

The university should also evaluate the cumulative effects of AI adoption. One AI application may appear low-risk when considered independently, but an institution may eventually use AI across admissions, teaching, assessment, student support, research, finance, human resources, and administration. The combined effect may be substantially different from the effect of each individual application.

This suggests that AI governance should operate at both system level and institutional level.

At the system level, the institution evaluates a particular AI application.

At the institutional level, it evaluates the combined portfolio of AI applications, shared vendors, common data sources, common infrastructure, overlapping risks, and cumulative effects on students and staff.

The institution should also establish mechanisms for periodic review because generative AI technology is developing rapidly. Capabilities, risks, pricing, vendors, models, regulatory requirements, and user practices may change over time.

An AI policy that is appropriate at one point may therefore require revision as technology and institutional experience develop.

The university should consequently treat AI evaluation as a continuing governance process rather than a one-time technology procurement exercise.

The higher-education case ultimately illustrates the broader AMTEF proposition that technology value is multidimensional. The success of generative AI in higher education cannot be established merely by showing that it produces fluent text quickly or reduces the time required for particular administrative tasks.

A comprehensive evaluation must ask whether it improves educational and institutional outcomes, whether it protects students and staff, whether it preserves academic standards, whether it supports rather than unnecessarily replaces human expertise, whether access is equitable, whether risks are governed, whether information is protected, and whether the institution can monitor and correct problems over time.

The final evaluation should therefore distinguish among capability, use, outcome, and consequence.

Capability asks:

What can the generative AI system do?

Use asks:

How is the institution actually using those capabilities?

Outcome asks:

What changes occur as a result of that use?

Consequence asks:

What wider effects occur for students, academics, administrators, the institution, society, and the environment?

This distinction is central to AMTEF because technological capability alone does not establish technological value.

The higher-education case therefore reinforces the central principle of this chapter: artificial intelligence should be evaluated not simply according to whether it functions, but according to whether it functions appropriately, produces meaningful value, manages foreseeable risks, respects affected stakeholders, and remains governable throughout its lifecycle.

10.28 AI Evaluation Protocol

Stage 1: define the use case.

Stage 2: identify affected stakeholders.

Stage 3: classify the potential level of harm.

Stage 4: evaluate data.

Stage 5: evaluate model performance.

Stage 6: assess fairness and bias.

Stage 7: assess security and privacy.

Stage 8: assess human oversight.

Stage 9: assess governance and accountability.

Stage 10: conduct controlled deployment.

Stage 11: monitor real-world performance.

Stage 12: reassess periodically.

10.29 Research Questions

How can AI systems be evaluated beyond technical accuracy?

How should fairness be incorporated into AI evaluation?

What forms of human oversight are effective?

How can organizations evaluate generative AI hallucination risk?

How does AI model drift affect long-term technology performance?

How should AI evaluation combine NIST, ISO, OECD, UNESCO and organizational frameworks?

10.30 Chapter Summary

AI evaluation must move beyond accuracy. Responsible evaluation considers reliability, fairness, transparency, explainability, privacy, security, human oversight, environmental impact, governance and social consequences.

Artificial intelligence is not simply a computational technology that produces predictions, classifications, recommendations, generated content, or automated actions. It is a socio-technical system embedded within organizational, economic, legal, cultural, environmental, and human contexts. Consequently, evaluating AI requires attention not only to what a model can technically achieve but also to how the technology is developed, deployed, used, governed, monitored, modified, and eventually retired.

The central argument of this chapter is that technical performance is necessary but insufficient for establishing technological value. An AI system may demonstrate high accuracy while producing inappropriate outcomes, operating on poor-quality data, creating privacy risks, reproducing historical inequalities, exposing organizations to cybersecurity threats, consuming substantial resources, or being deployed in a context for which it was never adequately evaluated.

AI evaluation should therefore begin with purpose. The evaluator should first establish what problem the AI system is intended to address, who experiences that problem, what alternatives exist, why AI is being considered, and what would happen if AI were not introduced. This prevents technology from becoming the starting point for a solution to a problem that has not been adequately defined.

The definition of the problem also determines the definition of success. An organization cannot evaluate whether an AI system is successful without establishing what successful performance means. Depending on the application, success may involve predictive accuracy, improved decision quality, reduced administrative workload, improved accessibility, faster service, reduced costs, better resource allocation, improved safety, improved learning outcomes, or other measurable outcomes.

Data constitute another fundamental dimension of AI evaluation. AI systems depend on data for training, validation, testing, operation, and monitoring. Data quality therefore directly influences system quality. Accuracy, completeness, consistency, relevance, timeliness, representativeness, provenance, integrity, and appropriate governance all require consideration.

The chapter has also demonstrated that more data do not automatically mean better data. Large datasets can contain systematic errors, missing information, historical inequalities, inappropriate labels, duplicated records, outdated information, or populations that are inadequately represented. The evaluator must therefore examine the quality and appropriateness of data rather than relying on data volume as an indicator of reliability.

The distinction between training, validation, and test data is also important. A model that performs exceptionally well on the data used for development may perform poorly when confronted with previously unseen cases. Overfitting, data leakage, inappropriate validation procedures, and unrealistic testing conditions can create an inflated impression of system capability. Evaluation must therefore focus on generalization to the environment in which the system will actually operate.

Accuracy, precision, recall, and F-measure provide useful technical indicators, but no single metric is universally appropriate. Metric selection should be connected to the purpose and consequences of the AI system. In some contexts, false positives may be particularly costly; in others false negatives may be more consequential. The evaluator must therefore understand the decision context before determining which performance measures are appropriate.

Fairness presents a further challenge because fairness is not represented by one universally applicable mathematical criterion. Different definitions of fairness may emphasize different relationships among groups, outcomes, errors, opportunities, or treatment. The evaluator must therefore identify the relevant populations, determine which fairness concerns are applicable, understand the distribution of outcomes and errors, and examine the social and institutional context in which the AI system operates.

Bias may enter at many stages of the AI lifecycle. It can arise during problem definition, data collection, sampling, measurement, labelling, feature selection, model development, threshold selection, deployment, human interpretation, and organizational use. Consequently, the evaluator should not search for bias only inside the algorithm.

The lifecycle perspective is therefore essential. The AI lifecycle begins with problem definition and continues through data collection, preparation, model development, training, validation, testing, deployment, monitoring, updating, and retirement. At every stage there are opportunities for quality problems, bias, risk, or unintended consequences to arise.

This means that AI evaluation should not be treated as a one-time activity undertaken immediately before deployment. A system that is appropriate at one point in time may become inappropriate as data distributions change, users change, organizational processes change, external conditions change, or new threats emerge.

Model drift provides an important illustration of this principle. Customer behaviour, economic conditions, disease patterns, fraud techniques, language, technology infrastructure, and other environmental conditions can change. A model that performed effectively when developed may therefore deteriorate when the environment changes.

Continuous monitoring is consequently an essential part of responsible AI evaluation. Monitoring should consider technical performance, fairness, data quality, security, availability, user complaints, unexpected outputs, drift, incidents, and other indicators relevant to the system's purpose.

The chapter has also emphasized the importance of explainability. An AI system may need to provide understandable reasons for its outputs, particularly when decisions have significant consequences for individuals. Explainability can support human review, debugging, accountability, appeals, regulatory assurance, and stakeholder understanding.

However, explainability should be proportionate to context. Not every AI system requires the same degree or form of explanation. A low-consequence recommendation system may require less explanation than an AI system used to support medical, financial, employment, educational, legal, or safety-related decisions.

Transparency is similarly broader than publication of source code. Meaningful transparency may involve information about data sources, model purpose, intended users, limitations, performance, risks, update procedures, governance arrangements, and system changes. The objective is to provide sufficient information for appropriate understanding and scrutiny while also protecting legitimate privacy, security, intellectual-property, and safety interests.

Generative AI introduces additional evaluation challenges. A system may generate highly fluent content that is factually inaccurate. Consequently, linguistic quality cannot be treated as a substitute for factual reliability. Generative AI should be evaluated for hallucination, citation accuracy, consistency, robustness, domain-specific reliability, harmful outputs, privacy leakage, inappropriate confidence, and the need for human verification.

The chapter has therefore emphasized the distinction between an AI system that produces plausible content and one that produces content that is sufficiently reliable for a particular purpose. The acceptable level of error depends on the consequences associated with the application.

Generative AI also introduces questions concerning intellectual property, academic integrity, workforce transformation, educational practice, privacy, cybersecurity, environmental resources, and institutional governance. These issues demonstrate why generative AI must be evaluated as an organizational and social technology rather than merely as a language or content-generation system.

AI agents introduce another level of complexity because they can move beyond generating information to performing sequences of actions. The evaluation of agents must therefore consider authorization, tool access, persistence, autonomy, escalation, error recovery, logging, human approval, external communication, financial transactions, and the possibility of irreversible consequences.

The distinction between generating a recommendation and executing an action is particularly important. An incorrect recommendation may be rejected by a human. An incorrect automated action may directly affect a person, organization, financial account, database, communication channel, or physical system.

As autonomy increases, therefore, the importance of boundaries, authorization, monitoring, human oversight, auditability, and recovery mechanisms also increases.

Human oversight has emerged as a central principle throughout the chapter. Human involvement should not be merely symbolic. A person who is formally responsible but routinely accepts every AI recommendation without meaningful review does not necessarily provide effective oversight.

Meaningful oversight requires understanding, authority, time, access to relevant information, the ability to challenge or override outputs, appropriate training, and clear accountability. The effectiveness of human oversight should itself be evaluated.

Privacy represents another major dimension. AI can create privacy risks through extensive data collection, inference, profiling, facial recognition, retention, model memorization, secondary use, and information sharing. Privacy evaluation must therefore consider both information explicitly collected and information that can potentially be inferred.

Cybersecurity is equally important. AI systems may be vulnerable to adversarial examples, data poisoning, model extraction, prompt injection, malicious inputs, unauthorized access, supply-chain compromise, and other attacks. Conventional cybersecurity remains necessary, but AI systems may require additional controls reflecting their distinctive architecture and behaviour.

Security should therefore be considered throughout the AI lifecycle, from data collection and model development to deployment, monitoring, updating, and retirement.

AI risk management provides a structured mechanism for addressing these challenges. NIST AI RMF organizes AI risk management around Govern, Map, Measure, and Manage. These functions complement rather than replace the broader AMTEF approach.

Govern establishes responsibility, authority, policies, accountability, and oversight.

Map establishes the context, stakeholders, intended purposes, risks, benefits, dependencies, and potential impacts.

Measure provides evidence concerning technical performance, fairness, security, privacy, reliability, and other relevant characteristics.

Manage connects identified risks to decisions, controls, mitigation, monitoring, escalation, and continuous improvement.

The chapter has also examined the relationship between specialized AI frameworks and the broader AMTEF architecture. AMTEF provides a multidimensional technology-evaluation framework within which AI-specific approaches can be incorporated. NIST AI RMF provides a structured AI risk-management approach. ISO/IEC 42001 provides an organizational AI management-system perspective. ISO/IEC 23894 provides guidance concerning AI-specific risk management. OECD and UNESCO provide broader principles and policy perspectives concerning responsible and human-centred AI.

These approaches should therefore be viewed as complementary layers rather than mutually exclusive frameworks.

AMTEF provides the broader evaluation architecture.

Specialized AI frameworks provide more detailed mechanisms for particular dimensions of that architecture.

This relationship is important because no single framework can fully capture every dimension of AI value and risk. Technical standards may provide detailed requirements for particular characteristics. Risk-management frameworks may provide structured processes for identifying and managing risks. Ethical frameworks may provide principles concerning human dignity, fairness, inclusion, and human oversight. Organizational management systems may establish responsibilities and processes. AMTEF provides a broader structure for bringing these dimensions together within an overall technology-evaluation process.

The chapter's case studies demonstrate the practical importance of this integrated approach.

The AI recruitment case illustrates how technical performance can coexist with concerns concerning historical bias, gender differences, socioeconomic effects, disability access, privacy, explainability, legal requirements, human review, appeals, and vendor dependency. The case demonstrates that an algorithm that accurately reproduces historical outcomes may not necessarily produce outcomes that are appropriate for future organizational decision-making.

The generative AI case in higher education illustrates a similarly multidimensional problem. Generative AI may support tutoring, administrative communication, academic writing, curriculum development, and research assistance while simultaneously creating questions concerning factual reliability, academic integrity, privacy, inequitable access, dependence, human interaction, intellectual property, cybersecurity, and institutional governance.

These cases demonstrate the difference between technological capability and technological value.

Capability asks:

What can the technology do?

Evaluation asks:

How well does it perform?

AMTEF goes further by asking:

What value does it create?

For whom does it create value?

What risks does it create?

Who bears those risks?

What unintended consequences may occur?

What governance mechanisms exist?

Can the system be monitored and corrected?

What happens when the system fails?

What happens when the environment changes?

What happens when the technology is no longer appropriate?

These questions transform AI evaluation from a narrow technical exercise into a comprehensive decision-support process.

The chapter has also established that AI evaluation should consider both intended and unintended consequences. Technology developers and organizations may design a system around a specific objective, but actual use can extend beyond the original design. Users may discover new applications, organizations may integrate the technology into additional processes, or systems may become connected to other technologies.

Consequently, the evaluator must consider function creep, secondary use, changing organizational practices, and evolving stakeholder expectations.

Risk should also be understood in relation to consequence and reversibility. A low-consequence and easily reversible error may be managed differently from an error that causes permanent or difficult-to-reverse consequences. This is particularly important for autonomous AI agents and AI systems involved in high-impact decision-making.

AI evaluation should therefore consider severity, likelihood, exposure, detectability, duration, reversibility, affected population, uncertainty, and available alternatives.

Environmental consequences should not be overlooked. AI systems require physical infrastructure, computing resources, electricity, cooling, data storage, networking, hardware manufacturing, maintenance, and eventual replacement or disposal. Environmental evaluation should therefore extend beyond the energy used during a particular model invocation and consider the broader lifecycle.

This connects AI evaluation with the environmental dimensions discussed elsewhere in the AMTEF framework.

Similarly, AI has economic consequences that extend beyond software licensing or computational costs. AI may change labour requirements, organizational structures, productivity, professional roles, training needs, service delivery, and competitive conditions. The economic evaluation should therefore consider both direct costs and wider organizational effects.

The human dimension is equally important. AI can change professional autonomy, decision-making authority, workload, skill requirements, responsibility, and relationships between humans and organizations. Employees and other users may experience AI as a tool for augmentation, a source of monitoring, a decision-support mechanism, or an automation mechanism. These different experiences can produce different outcomes.

Social consequences should also be considered. AI systems can influence access to employment, education, healthcare, financial services, information, public services, and other opportunities. The evaluator should therefore consider whether benefits and burdens are distributed appropriately across relevant populations.

Governance provides the mechanism for coordinating these dimensions. Effective AI governance should identify ownership, authority, responsibility, validation, monitoring, incident management, change approval, escalation, suspension, and retirement.

Accountability should be traceable. If a system produces a harmful or unexpected outcome, the organization should be able to determine what happened, which system version was involved, what data and configuration were used, who authorized the system, who operated it, what controls were active, and what corrective actions are available.

Auditability is therefore closely connected with governance. Logs, model versions, data documentation, decision records, access records, change histories, performance records, and incident reports provide evidence that can support investigation and accountability.

Incident management completes the lifecycle. AI evaluation does not end when the system is deployed. When an incident occurs, the organization should detect it, report it, classify it, contain it, investigate it, communicate appropriately, remediate the problem, recover operations, learn from the event, and reassess the technology.

The resulting lifecycle can be represented as:

Identify → assess → design controls → validate → deploy → monitor → detect → respond → recover → learn → redesign → reassess.

This continuous cycle is central to responsible AI evaluation.

The chapter therefore reinforces a fundamental AMTEF principle: no single dimension can adequately describe technological value.

Technical performance cannot by itself establish social value.

Economic efficiency cannot by itself establish ethical acceptability.

Regulatory compliance cannot by itself establish responsible use.

Human oversight cannot by itself guarantee fairness.

Transparency cannot by itself guarantee reliability.

High accuracy cannot by itself guarantee appropriate outcomes.

And innovation cannot by itself establish that a technology should be deployed.

A multidimensional evaluation must bring these dimensions together while recognizing that different technologies and applications will require different levels of attention.

The ultimate purpose of AI evaluation is therefore not to produce a single universal score for artificial intelligence. It is to provide a structured evidence base for informed technological decisions.

The decision may involve adoption, modification, restricted deployment, additional testing, continued monitoring, expansion, suspension, replacement, or retirement. The appropriate decision depends on the evidence, purpose, context, consequences, alternatives, governance arrangements, and risk tolerance associated with the specific application.

AMTEF provides the broader architecture for making such evaluations systematic. Within that architecture, specialized AI frameworks can provide detailed methods for managing particular dimensions of AI risk and governance.

(CONCEPTUAL) FRAMEWORK

AI Capability to Evidence-Based Decision and Continuous Reassessment

Figure: Conceptual Framework for Multidimensional AI Evaluation

The framework conceptualizes AI evaluation as a continuous and multidimensional process. AI capability enables AI application, which is progressively evaluated across technical, human, ethical, legal, privacy, cybersecurity, environmental, economic, risk, resilience, governance, and accountability dimensions. The resulting evidence informs decision-making, which feeds into continuous learning, monitoring, and reassessment throughout the AI lifecycle.

The central conclusion of this chapter is therefore that artificial intelligence should be evaluated as a dynamic socio-technical system rather than as a static algorithm.

Its quality depends not only on the model but also on the data, problem definition, application, users, organization, governance system, infrastructure, environment, and wider social context.

AI evaluation must consequently remain multidimensional, lifecycle-oriented, evidence-based, context-sensitive, and proportionate to consequence.

The most important question is no longer simply:

Does the AI system work?

The broader evaluation questions are:

Does it work for the intended purpose?

Does it work reliably in the environment in which it is used?

Does it work appropriately for different affected populations?

Can its limitations be understood?

Can important decisions be challenged?

Are privacy and security adequately protected?

Is human oversight meaningful?

Are risks governed throughout the lifecycle?

Are benefits and burdens appropriately considered?

Can the organization detect and respond when the system fails?

Can the system be modified, restricted, or retired when circumstances change?

These questions establish the foundation for the next stage of the AMTEF framework, in which multidimensional technology evaluation can be translated into systematic research questions, evidence requirements, assessment procedures, decision rules, and practical organizational evaluation processes.

10.31 Discussion Questions

Why can a highly accurate AI system still be unacceptable?

What is the difference between model performance and social impact?

How should human oversight be evaluated?

What makes generative AI different from traditional predictive AI?

What are the risks of treating AI governance as a purely technical activity?