Governance Meaning and Problem Context

Defines governance in RAIDT as the practical arrangement of oversight, control, accountability, reviewability, contestability, reconstructability, and continuous improvement around each run of a generative AI system. This star anchors governance in observable organisational practice rather than in abstract ethical aspiration.

This is a foundations note with strong links to theory, evidence design, implementation logic, and policy translation. It explains why RAIDT is framed as a governance project rather than only a Responsible AI or technical assurance project. It helps position RAIDT across several layers at once: conceptual foundations, run-level evidence design, five-pillar scoring, empirical validation, governance interventions, and policy alignment. For supervisors, it clarifies the problem RAIDT is trying to solve. For workshops, it provides the shared vocabulary needed before discussing evidence packs, scoring, RAG, prompt engineering, or sector playbooks.

• What does governance mean in the context of generative AI used in organisational work?
• Why does RAIDT need a governance concept that is more precise than a broad Responsible AI label?
• What organisational problem is being solved when RAIDT treats the run as the unit of governance?
• Why are oversight, control, accountability, reviewability, contestability, reconstructability, and continuous improvement central governance functions?
• What kinds of failure or uncertainty make run-level governance necessary?
• What evidence shows whether governance is actually being performed rather than merely claimed?
• How does governance connect to the run-level evidence pack?
• How does governance connect to the five RAIDT pillars of Responsibility, Auditability, Interpretability, Dependability, and Traceability?
• How does this star help supervisors understand the conceptual scope of the project?
• How does this note support Paper 08, Paper 09, Paper 10, and sector playbooks?

• 10–20 min - What is lost when governance is treated as a general ethics slogan rather than a set of operational practices around a specific AI run?
• 20–40 min - Why did the RAIDT project emerge from Responsible AI, managerial uncertainty, and governance pressure rather than from model performance alone?
• 40–60 min - If the run is the unit of governance, what evidence must an organisation capture to defend, review, contest, and improve a generative AI-assisted decision or work product?

Governance is often invoked in AI discussions as a broad signal of seriousness, responsibility, or compliance. In many organisational settings, however, the term remains underspecified. Teams may claim to govern AI because they have a policy, an ethics principle, a model approval checklist, or a vendor assurance statement. Those instruments matter, but they do not by themselves explain how a specific use of a generative AI system is overseen, checked, reconstructed, or challenged when it affects real work. RAIDT starts from that gap. It treats governance not as an aspirational label but as an operational problem: how an organisation can responsibly direct, observe, justify, and improve the use of generative AI in situated work tasks under conditions of uncertainty.

The key idea in this star is that governance should be understood through the run. A run is one configured use of a generative AI system for a specific task, at a specific time, in a specific context. It includes the instruction or prompt, model choice, tool configuration, retrieved context where relevant, output, and the human or automated checks applied before the result is used. This matters because many governance failures in generative AI are not visible at the level of abstract policy or even at the level of the model alone. They arise in the interaction between configuration, context, task framing, retrieval, output handling, and downstream use. A system may be approved in principle but still produce an unreviewable, poorly evidenced, or unsafe run in practice.

RAIDT therefore interprets governance as the organised capacity to oversee runs, shape them through controls, assign accountability, review outcomes, allow contestation where necessary, reconstruct what happened, and improve future practice. Oversight means there is an identifiable authority or process that can examine how generative AI is being used. Control means there are design choices, guardrails, prompts, retrieval settings, approval steps, and escalation routes that influence behaviour before harm occurs. Accountability means named actors can justify why a run was initiated, how it was checked, and who is responsible for the final use of the output. Reviewability means another competent party can inspect the run and assess whether the process and outcome were acceptable. Contestability means affected actors can question or challenge consequential outputs. Reconstructability means the organisation can piece together what happened in enough detail to investigate problems. Continuous improvement means lessons from runs feed back into prompts, workflows, tool settings, training, governance rules, or escalation criteria.

This framing addresses a specific problem context. Generative AI introduces a combination of technical and managerial uncertainty. Technical uncertainty arises because outputs can vary, retrieved context can be partial or misleading, alignment controls such as RLHF do not guarantee suitability for a local organisational task, and lightweight adaptation methods such as PEFT or LoRA can alter behaviour in ways that are beneficial yet difficult to interpret. Managerial uncertainty arises because organisations must decide when use is legitimate, what evidence is sufficient, which checks are proportionate, how responsibility is shared between human and system, and how to satisfy internal governance as well as external expectations such as the EU AI Act, ISO/IEC 42001, or the NIST AI RMF. In other words, organisations do not only face the risk of model error. They face the risk of not being able to explain, defend, or improve AI-assisted work.

RAIDT responds by making governance empirically traceable. Its two practical outputs are the run-level evidence pack and the five-pillar RAIDT score profile. The evidence pack captures the materials needed to examine a run: task purpose, prompt or instruction, model and tool settings, retrieval details in RAG workflows, output, checks, reviewer input, exceptions, and final disposition. The score profile then assesses the run across Responsibility, Auditability, Interpretability, Dependability, and Traceability. Governance is therefore not external to RAIDT. Governance is the reason the framework exists and the logic that ties evidence collection to evaluative scoring.

This also clarifies why RAIDT is not reducible to conventional Responsible AI discourse. Responsible AI often provides normative principles such as fairness, transparency, accountability, safety, and human oversight. Those principles remain valuable, but organisations still need a unit of analysis small enough to govern concrete use and rich enough to capture context. RAIDT proposes that the run can serve that function. It translates high-level principles into observable governance practice. For example, traceability is strengthened when prompts, retrieved documents, versioned tools, and reviewer actions are recorded. Auditability is strengthened when the sequence of decisions around a run can be inspected by an internal auditor or supervisor. Responsibility is strengthened when a named owner authorises use and a named reviewer checks suitability before release.

Practical examples make the point clearer. In a university setting, a generative AI tool may draft a student-support response based on policy documents retrieved through RAG. Governance questions arise immediately: which documents were retrieved, was the prompt constrained, who checked the answer, and can the institution reconstruct the run if a student challenges the advice? In a healthcare administration setting, a model may summarise referral notes. Even if it performs well on average, governance still requires role-based oversight, review thresholds, and evidence showing what source material was used. In a consulting firm, a customised model adapted through LoRA may generate client-facing analysis. Governance is not satisfied by noting that the model was fine-tuned; the organisation must also evidence which configuration was active in the run, who approved the workflow, and how quality or risk checks were performed.

This star matters because it prevents governance from drifting into vagueness. Without a precise meaning, governance can collapse into one of three weak substitutes: ethics language without operational consequence, compliance language without evidential depth, or technical assurance language without organisational accountability. RAIDT requires a stronger position. Governance must be inspectable in use. That is why this star sits early in the project architecture. It defines the governing problem that later stars, items, evidence templates, scoring rules, empirical studies, and policy pathways must answer.

There are, however, important boundaries. Governance in RAIDT does not claim to eliminate uncertainty, remove the need for human judgement, or fully solve all normative disputes around AI use. It does not claim that every run needs the same degree of evidence or review. Proportionality still matters. Nor does it claim that run-level evidence replaces wider organisational governance, procurement, security, data protection, or model risk management. Instead, it complements those structures by providing a practical unit at which oversight and learning can occur.

For the overall project, this star supplies the conceptual bridge between theory and implementation. It explains why the run-level evidence pack is necessary, why the five pillars are governance-relevant rather than merely descriptive, why Paper 08 can frame RAIDT as a methodological contribution, why Paper 09 can empirically validate run-level governance claims, and why Paper 10 can translate the framework into policy and sector playbooks. In short, this star explains the meaning of governance and the problem context that makes RAIDT necessary.

Q1. What does governance mean in RAIDT?

Answer:
In RAIDT, governance means the practical arrangement through which an organisation directs, checks, justifies, and improves the use of generative AI in real work. It includes oversight, control, accountability, reviewability, contestability, reconstructability, and continuous improvement. The definition is deliberately operational so that governance can be evidenced rather than merely declared.

Practical example:
A policy team uses a GenAI assistant to draft a consultation summary. Governance exists only if the task owner, prompt setup, retrieved materials, review process, and approval decision are clear and inspectable.

Link to RAIDT:
This definition underpins the run-level evidence pack and justifies why RAIDT scores a run across the five pillars rather than relying on generic organisational claims.

Q2. Why is a broad Responsible AI label not enough?

Answer:
Responsible AI principles are important, but they often remain high-level and do not specify what must be captured, checked, or reviewed in a concrete organisational use. RAIDT narrows the question from "Is this AI responsible?" to "Was this run governed in a way that can be evidenced and defended?"

Practical example:
An organisation may state that it values transparency, yet if it cannot show which prompt and documents produced a given output, the transparency claim is weak in practice.

Link to RAIDT:
RAIDT translates principle language into run-level evidence requirements and measurable pillar scores.

Q3. Why does RAIDT treat the run as the unit of governance?

Answer:
A run is where the task, context, model configuration, retrieval behaviour, output, and checking process come together. Governance failures often emerge in that situated combination rather than at policy level alone. Using the run as the unit allows governance to be specific, reconstructable, and proportionate.

Practical example:
Two teams may use the same model, but one uses it for low-risk brainstorming while the other uses it to draft regulatory advice. The governance needs differ by run, not just by model.

Link to RAIDT:
The run is the basis of the evidence pack and the object being scored across Responsibility, Auditability, Interpretability, Dependability, and Traceability.

Q4. What problem is this star trying to solve?

Answer:
This star addresses the ambiguity surrounding AI governance in organisations. Many institutions know they should govern GenAI but lack a clear operational definition of what governance requires in practice. RAIDT solves this by specifying governance as evidence-bearing activity around runs.

Practical example:
A supervisor asks why a chatbot-generated recommendation should be trusted. Without run-level governance, the team can only offer general statements. With RAIDT, it can show the run record, checks performed, and score profile.

Link to RAIDT:
The star gives conceptual legitimacy to evidence capture, scoring, and governance interventions.

Q5. How does governance relate to uncertainty?

Answer:
Governance becomes essential because generative AI introduces uncertainty in outputs, context use, model behaviour, and downstream interpretation. Managerial uncertainty is equally important: organisations must decide who may rely on outputs, what evidence is enough, and how exceptions are handled.

Practical example:
A RAG-based internal assistant retrieves an outdated policy memo. The problem is not only retrieval error; it is also whether the workflow required source review before use.

Link to RAIDT:
RAIDT manages uncertainty through evidence capture, proportionate checks, and pillar-based assessment of whether the run was governable.

Q6. What evidence shows that governance actually happened?

Answer:
Useful evidence includes the task purpose, user role, prompt or instruction, model and tool configuration, retrieved sources, timestamps, output, review notes, escalation decisions, exception handling, and final approval or rejection. Governance is evidenced by the presence and coherence of these materials.

Practical example:
If a manager disputes an AI-generated summary, the team should be able to show the prompt, source documents, reviewer comments, and reason the summary was accepted or revised.

Link to RAIDT:
These are core inputs to the run-level evidence pack and support scoring, audit, and intervention design.

Q7. How do the five RAIDT pillars express governance?

Answer:
Each pillar reflects one governance dimension. Responsibility asks who owns and answers for the run. Auditability asks whether the run can be inspected. Interpretability asks whether the logic of the run can be understood well enough for its purpose. Dependability asks whether the run is reliable enough in context. Traceability asks whether the chain from input and context to output and decision can be followed.

Practical example:
A run may be dependable in producing fluent text but weak in traceability if its retrieved sources are not recorded.

Link to RAIDT:
The five-pillar profile is the evaluative expression of governance quality at run level.

Q8. Why are contestability and reconstructability important?

Answer:
Contestability allows stakeholders to question consequential AI-assisted outputs. Reconstructability allows investigators or reviewers to determine what happened and why. Without both, governance becomes ceremonial because problematic runs cannot be challenged or examined properly.

Practical example:
An employee receives an AI-assisted performance summary they believe is misleading. Governance requires a route to challenge it and sufficient records to reconstruct the run.

Link to RAIDT:
Contestability and reconstructability shape evidence requirements and support stronger Auditability and Traceability scores.

Q9. How does this note relate to prompt engineering, RAG, and adaptation methods such as PEFT or LoRA?

Answer:
These technical practices are not separate from governance. Prompt design influences control. RAG influences traceability and reviewability because retrieved context affects the output. PEFT or LoRA influence reconstructability and interpretability because configuration changes must be known if behaviour is to be explained.

Practical example:
A fine-tuned internal model produces a different answer from the base model. Governance requires the run record to show which adapted configuration was active and whether additional review was required.

Link to RAIDT:
RAIDT turns technical configuration choices into governed evidence fields within the run pack.

Q10. How does this star help supervisors understand the project?

Answer:
It gives a clear answer to the question, "What exactly is the governance problem that RAIDT solves?" By defining governance precisely, the project can show how its methods, scoring logic, empirical validation, and policy implications all follow from a common problem statement.

Practical example:
In supervision, the note can be used to explain why the thesis is not merely about AI ethics or auditing in general, but about governing situated GenAI use through run-level evidence.

Link to RAIDT:
This star is the conceptual entry point for Paper 08, the justification for measurement in Paper 09, and the basis for policy translation in Paper 10.

• A university admissions office uses GenAI to draft applicant communication. Governance requires evidence of the prompt, policy sources used, staff review, and the point at which a human approves or rejects the draft.
• A hospital administration team uses a summarisation assistant to prepare case-handling notes. Governance requires role-based controls, source visibility, reviewer sign-off, and a record of whether the output entered a consequential workflow.
• A public sector unit uses a RAG system to answer staff queries on procurement rules. Governance requires the retrieval set, versioned policy documents, prompt constraints, and evidence of correction when the answer is incomplete or wrong.
• A consulting practice deploys a LoRA-adapted model for internal report drafting. Governance requires clear configuration records, assurance about permitted uses, quality review criteria, and the ability to reconstruct the run if a client challenges the analysis.
• A regulated firm uses GenAI to support compliance monitoring. Governance requires escalation thresholds, exception logging, reviewer identity, and evidence that the AI-assisted output did not bypass mandated human judgement.

• Run identifier and timestamp
• Task purpose and organisational context
• User role and accountable owner
• Prompt or instruction text
• Model name, version, and relevant configuration
• Tooling used, including plugins, agents, or external systems
• Retrieval details for RAG, including documents, versions, and query logic where feasible
• Adaptation state, such as PEFT or LoRA configuration where relevant
• Alignment or safety controls in force, including RLHF-derived system behaviour where relevant to the use case
• Output produced by the run
• Human review steps, reviewer identity, and review criteria
• Automated checks, filters, or validation tools applied
• Exceptions, overrides, escalation decisions, and rationale
• Final disposition: accepted, revised, rejected, or deferred
• Link to score profile across the five RAIDT pillars
• Notes for improvement, retraining, prompt revision, or workflow redesign

• S2.01 · Governance meaning
• S2.02 · Oversight
• S2.03 · Control
• S2.04 · Accountability
• S2.05 · Reviewability
• S2.06 · Continuous improvement
• S2.07 · Contestability
• S2.08 · Reconstructability
• S2.09 · Organisational legitimacy
• S2.10 · GenAI failure modes

This star connects directly to Paper 08: foundations and methodological pathways because it defines the governance problem that justifies RAIDT's run-level unit of analysis and its evidence-based methodological design. It supports Paper 09: empirical validation by clarifying what must be observed, measured, and compared in real organisational runs if the framework is to show practical validity. It underpins Paper 10: policy pathways by showing how operational governance can be translated into defensible alignment with policy and standards rather than remaining at the level of abstract principle.

It also supports sector playbooks because different domains need concrete guidance on how governance expectations become run-level evidence requirements, thresholds, and interventions. The note links directly to RAIDT scoring by explaining why the five pillars are governance dimensions rather than isolated metrics. It links to the RAIDT evidence pack because governance is only inspectable when run materials are captured systematically. Finally, it links to RAIDT governance interventions because weak scores or repeated failures should trigger practical changes such as prompt redesign, reviewer escalation, workflow constraints, additional training, or revised approval rules.

• Responsible AI governance literature
• Information Systems governance and accountability literature
• Organisational uncertainty and managerial decision-making literature
• AI auditability, traceability, and contestability literature
• Human oversight and reviewability guidance in AI governance
• Generative AI risk and failure mode literature
• RAG documentation and evidence-chain discussions
• PEFT / LoRA and model adaptation literature where governance implications are discussed
• RLHF, alignment controls, and limitations literature
• Policy and standards materials related to the EU AI Act, ISO/IEC 42001, and NIST AI RMF
• Empirical studies on human-AI collaboration, review practices, and assurance workflows

This note does not claim that governance can remove all uncertainty from generative AI use. It does not claim that run-level evidence alone is sufficient for legal compliance, information security, or enterprise-wide model risk management. It does not imply that every run requires maximal documentation; proportionality remains necessary. It does not claim that the five RAIDT pillars capture every possible normative concern. It does not replace deeper evaluation of fairness, privacy, safety, or sector-specific regulation where those issues are material. Instead, the note defines a practical governance lens that makes GenAI use more reviewable, contestable, and improvable in organisational settings.

This star explains what governance means in the RAIDT project and why that meaning matters. My argument is that governance should not be treated as a vague ethics term or as a purely top-level policy concern. In generative AI, the real governance problem appears in the run: one situated use of a model for a particular task, with a particular prompt, configuration, context, output, and checking process. RAIDT therefore treats the run as the unit of governance. That allows us to ask concrete questions: who authorised the run, what controls shaped it, what evidence was captured, how could it be reviewed, and how could it be challenged or improved? This star is foundational because it explains why the framework needs a run-level evidence pack and why the five pillars are governance-relevant. It also connects the project to Responsible AI, uncertainty, Information Systems governance, and policy pathways such as the EU AI Act, ISO/IEC 42001, and NIST AI RMF. In supervisory terms, this note defines the problem RAIDT solves.

• Why governance needs a sharper meaning in GenAI
• Why RAIDT uses the run as the unit of governance
• What governance functions must exist around a run
• What evidence must be captured
• How the five pillars operationalise governance
• What kinds of real organisational problems this solves
• Where the boundaries are
• Why this star matters for Papers 08, 09, and 10