Appendix B: Glossary of Key Terms
This glossary defines terms as they are used throughout this book. The definitions are intentionally practical, governance-oriented, and context-specific rather than technical or vendor-dependent. They reflect the central perspective developed throughout the book:
Generative AI systems may assist analytical work, but responsibility for evidence, interpretation, reasoning, and accountability remains human. :contentReferenceoaicite:0
8.6 AI and analytical workflow
8.6.1 Assistant
In this book, an assistant refers to the role assigned to a generative AI system within analytical workflows. An assistant may support inquiry, drafting, organization, comparison, or reflection, but it does not determine analytical direction, validate conclusions, or assume responsibility for outcomes.
8.6.2 Generative AI system
A generative AI system is a software system that produces text or other outputs by generating statistically likely continuations based on patterns learned from data. Throughout this book, these systems are treated as tools that may support analytical work, but not as sources of expertise, understanding, or authority.
8.6.3 Prompt (as a research instrument)
A prompt is a structured input used to guide interaction with a generative AI system. Similar to other research instruments, prompts shape inquiry by framing tasks, constraining outputs, embedding assumptions, and influencing what becomes analytically visible. Prompt design may influence exploration and interpretation, but it does not make generated outputs valid or authoritative.
8.6.4 AI-assisted analysis
AI-assisted analysis refers to analytical work in which generative AI contributes to activities such as drafting, synthesis, organization, comparison, or exploratory inquiry while responsibility for evidence, interpretation, reasoning, and conclusions remains human.
8.7 Governance and accountability
8.7.1 Governance boundary
A governance boundary is the conceptual and operational limit defining how generative AI may appropriately participate in analytical work. Governance boundaries distinguish between acceptable forms of assistance, conditional or high-review uses, and non-delegable human responsibilities such as validation, authoritative interpretation, and decision-making.
8.7.2 Conditional or high-review use
Conditional or high-review use refers to AI-assisted analytical activity that may be appropriate only when accompanied by additional safeguards such as evidence traceability, uncertainty disclosure, peer review, documentation, or formal sign-off because of its interpretive or institutional impact.
8.7.3 Accountability
Accountability refers to responsibility for analytical framing, interpretation, reasoning, evidence use, uncertainty communication, and conclusions. In AI-assisted work, accountability always remains human regardless of the extent of AI involvement.
8.7.4 Analytical integrity
Analytical integrity refers to the preservation of evidentiary discipline, methodological rigor, transparency, uncertainty awareness, and accountable reasoning throughout analytical workflows.
8.7.5 Disclosure
Disclosure is the practice of communicating how generative AI assistance was used within an analytical workflow. Disclosure should remain proportional to the role AI played and should support transparency, reviewability, and accountability.
8.7.6 Review discipline
Review discipline refers to the structured process of verifying, questioning, revising, and evaluating AI-assisted outputs before they are accepted into analytical work. This includes evidence checking, uncertainty assessment, reasoning evaluation, governance review, and accountability review.
8.8 Evidence and reasoning
8.8.1 Evidence vs. interpretation
This distinction separates source material such as data, documents, or empirical findings from the analytical judgments drawn from them. Generative AI may assist with summarizing or reorganizing evidence, but interpretation and justification remain human responsibilities.
8.8.2 Evidence traceability
Evidence traceability is the ability to connect analytical claims, summaries, interpretations, or recommendations back to identifiable and reviewable evidence sources.
8.8.3 Traceability
Traceability refers more broadly to the ability to connect analytical claims back to their sources, assumptions, reasoning processes, or explicit judgments. Maintaining traceability helps ensure that AI-assisted outputs remain interpretable, reviewable, and accountable.
8.8.4 Traceability ladder
The traceability ladder is a conceptual model showing how analytical work moves from primary evidence toward synthesis, interpretation, and decision-support language. As analytical distance from source material increases, uncertainty and review requirements also increase.
8.8.5 Verification
Verification is the process of checking factual claims against primary sources, authoritative references, or validated evidence. Verification is distinct from validation because generative AI outputs require human verification and cannot validate themselves.
8.8.6 Validation
Validation is the process of confirming that analytical conclusions, methods, or outputs are appropriate, defensible, and institutionally acceptable. In this book, validation remains a human responsibility and cannot be delegated to generative AI systems.
8.9 Failure modes and safeguards
8.9.1 Failure mode
A failure mode is a recurring pattern through which AI-assisted analysis can go wrong, often subtly. Failure modes discussed throughout this book include fluency masking uncertainty, framing lock-in, false balance, invisible assumptions, automation complacency, and diffusion of responsibility.
8.9.2 Fluency masking uncertainty
This failure mode occurs when coherent and confident AI-generated prose obscures evidentiary weakness, ambiguity, incomplete reasoning, or unresolved assumptions.
8.9.3 Framing lock-in
Framing lock-in occurs when early AI-generated framing choices narrow subsequent analysis by shaping which questions appear important, which interpretations remain visible, or which alternatives are considered legitimate.
8.9.4 False balance
False balance occurs when competing positions are presented as equally credible despite substantial differences in evidence quality, methodological rigor, or empirical support.
8.9.5 Invisible assumptions
Invisible assumptions are assumptions embedded within prompts, institutional norms, training data, or analytical conventions that remain unstated yet still shape interpretation and conclusions.
8.9.6 Diffusion of responsibility
Diffusion of responsibility is a failure mode in which ownership of analytical choices becomes blurred because AI-generated outputs are implicitly treated as responsible for wording, framing, or conclusions.
8.9.7 Automation complacency
Automation complacency refers to the gradual reduction in review intensity or analytical skepticism caused by repeated exposure to fluent and plausible AI-generated outputs.
8.9.8 Mitigation practice
A mitigation practice is a review or governance strategy intended to reduce the likelihood or impact of analytical failure modes. Mitigation practices discussed throughout this book include evidence verification, explicit uncertainty review, assumption checks, peer review, and accountable sign-off.
8.10 Core governing principle
8.10.1 “Generative AI assists; analysts decide.”
This statement represents the central governing principle of the book. Generative AI systems may support drafting, organization, synthesis, comparison, and exploratory inquiry, but responsibility for evidence, interpretation, reasoning, conclusions, and analytical impact remains human.