| 🏠 Back to Exam Syllabus | 📺 RooCloud on YouTube | 🌐 RooCloud Practice Exams |
AI Agency: Logging, Observability, Human in the Loop, and Hallucinations
This episode of the ISACA Advanced in AI Audit (AAIA) exam prep series explores what happens when AI systems begin making decisions on their own and the supervision techniques used to keep them under control. You’ll see how teams trace independent machine choices, where human checkpoints belong, and how to recognize and contain the failure modes unique to autonomous systems. The discussion equips auditors to judge whether a proposed AI tool can be safely deployed in a given business workflow.
What this episode covers
- What AI agency is — independence in decision-making and the four headaches it creates around accountability, self-regulation, values, and explainability.
- Logging and the chain of thoughts — capturing inputs, model parameters and versions, and outputs without drowning in nano decisions.
- The four categories of observability — data pipelines, infrastructure and system health, model performance, and interpretability.
- Three interpretability strategies — counterfactual explanations, feature visualization, and influential instance analysis.
- Human in the loop (HITL) — the high-stakes safety net that pauses a workflow for explicit human approval.
- Hallucinations — what they are, their three main causes, and how guardrails and templates contain them.
- Testing the surrounding software — why standard testing still applies to the non-AI components around the model.
Watch the full episode above for the worked examples and detailed explanations of each concept.
Frequently Asked Questions
What is AI agency and what risks does it create?
Agency is the ability of a machine to act independently, taking action on its own without waiting for a human to push a button. It creates four practical headaches for auditors: accountability for who is legally responsible when something goes wrong, self-regulation rules the machine follows, conflicting values when an efficient decision ignores human empathy, and the expectation of explainability when even the programmers cannot untangle the math.
What are the four categories of AI observability?
The four categories are data pipelines, which monitor data quality and catch malicious actions like prompt injection; infrastructure and system health, which tracks processing power and storage; model performance, which compares inputs to outputs and user feedback to detect model drift; and interpretability, which uses counterfactual explanations, feature visualization, and influential instance analysis to peek inside the black box.
What is the human in the loop strategy?
Human in the loop is the ultimate safety net for highly critical decisions, where the AI does not act alone. The workflow pauses and a human must give explicit approval before the action is finalized, like an autopilot that flies the plane while a human pilot monitors and can take over. Because it intentionally slows the system, it is reserved for situations where the risk is extremely high.
What causes AI hallucinations and how are they controlled?
A hallucination is when the software confidently presents an answer disconnected from reality. The three main culprits are flawed training data, a lack of real-world grounding, and ambiguity in the data input. To fight them, teams use strict guardrails on what is allowed in and out of the system and templates that force answers into a rigid, consistent structure with less room to wander into fantasy.
📚 Master the ISACA AAIA Exam!
Ready to test your knowledge? Access chapter-specific Multiple Choice Questions (MCQs) and full-length practice exams for the ISACA AAIA certification at RooCloud.com. Solve the chapter-wise questions to reinforce this lesson before moving to the next episode.
Reference: This article is based on concepts discussed in AI Agency: Logging, Observability, HITL & Hallucinations.