• qa_deck_100#slide 24 · Uncertainty, hallucination, and over-reliance

Over-reliance remains a material risk even when a human is formally in the loop because human presence does not by itself guarantee meaningful scrutiny. The managerial decision-making paper identifies overconfidence in opaque outputs and a disconnect between technical predictions and human judgement as central weaknesses of AI in uncertain settings. The evidence-review paper adds that harms emerge through workflow integration, staff incentives, over-reliance, and the routine use of outputs in downstream decisions. Together, these arguments show that a human reviewer may simply ratify a fluent recommendation, especially under time pressure, ambiguous evidence, or organisational expectations to move quickly.

For RAIDT, the issue is therefore not whether a person clicked approve, but whether the human role operated as a genuine governance control. Hybrid and assemblage models require explicit decision boundaries, uncertainty communication, and override capability; otherwise the 'human in the loop' becomes a symbolic safeguard rather than an effective one. This is where influence methods as governance interventions matter: the governance design must shape user behaviour towards challenge, escalation and documented judgement rather than passive acceptance. A run-level evidence pack is important because it records review decisions, edits, escalations and approvals, allowing the score profile to distinguish substantive oversight from nominal sign-off. If the evidence only shows a confident output and a rapid approval, Responsibility and Auditability remain weak even though a human was technically present.

Consider an infrastructure manager using RAIDT-style support for resource allocation under conflicting forecasts. The system produces a confident ranking of maintenance priorities after combining sensor feeds, narrative reports and retrieved policy material. Because the deadline is tight, the manager accepts the recommendation with minimal challenge and forwards it into budgeting discussions. A human was involved, but the human functioned mainly as a relay.

If later review shows that the recommendation rested on ambiguous inputs or poorly grounded assumptions, the problem is over-reliance, not the absence of a human. Governance depends on whether the run-level evidence pack shows uncertainty cues, source checks, edits, or escalation actions. Mere approval is not equivalent to meaningful oversight.

• 01-Responsible_AI_for_Managerial_Decision-Making_Under_Uncertainty-V3
• 13-RAIDT-Evidence-Review_M_v10
• 15-RAIDT-IS-Governance_M_v07