• qa_deck_100#slide 23 · Uncertainty, hallucination, and over-reliance

Hallucination is a governance problem because the organisational issue is not only whether a generated claim is factually wrong, but whether the organisation can later justify, reconstruct and contest how that claim entered a consequential workflow. The RAIDT papers argue that generative-AI risk materialises at run time: prompts, retrieved material, tools, provider versions, settings and review actions all shape one output in context. A hallucinated statement therefore exposes a failure of reviewability and accountability, not merely a lapse in technical accuracy. If a disputed output cannot be traced back to the configured use that produced it, policy documents and model-level descriptions remain too coarse to answer the core governance question: what happened in this case.

This is why RAIDT treats run as the unit of governance and uses a run-level evidence pack rather than relying on abstract assurances. Hallucination cuts across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). It weakens Responsibility when human oversight is nominal; Auditability when the run cannot be reconstructed; Interpretability when users cannot see the evidential basis of a claim; Traceability when prompts, sources and tool calls are not retained; and Dependability when the same task produces unstable or weakly grounded outputs. In practice, a strong score profile is not a guarantee of truth, but a low score profile, especially at anchors 1=missing / 3=partial / 5=audit-ready, signals that the organisation lacks the evidence needed to govern the risk. Hallucination therefore matters because it turns a content error into a governance failure with implications for audit, dispute resolution and organisational learning.

In the HR example used in the evidence-review paper, a manager uses GenAI to draft a performance appraisal. Suppose the draft states that an employee breached a policy clause or missed a target that was not supported by the retrieved HR policy text or the underlying appraisal notes. That is a hallucination, but it is also a governance issue because the draft may become part of an employment record and could be challenged by the employee.

A RAIDT response would not stop at correcting the sentence. Reviewers would inspect the run-level evidence pack: prompt template version, retrieved policy text, model and deployment settings, generated output hash, and the manager's review actions. Without that evidence, the organisation cannot tell whether the problem came from unsupported generation, stale retrieval, hidden configuration change, or inadequate human checking.

• 13-RAIDT-Evidence-Review_M_v10
• 15-RAIDT-IS-Governance_M_v07