🎯 Learning Objectives
- Identify bias sources in AI outputs and mitigate them
- Protect sensitive data in AI workflows
- Navigate regulatory and compliance risks in AI use
- Design human-in-the-loop systems
- Build prompt governance frameworks for organisations
1. Bias & Model Limitations
Common Bias Categories
- Recency bias — over-weighting recent training data
- Geographic bias — models trained predominantly on English-language, Western sources
- Confirmation bias — models tend to validate the framing of the question
- Availability bias — common information produces more confident outputs regardless of accuracy
Anti-bias instruction set (add to any analytical prompt):
After generating your response:
1. Identify the assumptions embedded in your answer
2. Name which perspective your answer implicitly favours
3. Provide one counterpoint from a different cultural/regional perspective
4. Flag any conclusion presented as universal that may not be
2. Data Sensitivity & Confidentiality
| Data Category | AI Usage | Mitigation if Used |
|---|---|---|
| Public information | Unrestricted | None required |
| Internal business data | Use with caution | Anonymise specifics; check vendor policy |
| Personal data (PII) | Avoid or anonymise | Replace with [NAME], [COMPANY], [DATE] |
| Commercially sensitive | Avoid unless private instance | On-premise deployments only |
| Regulated (health, financial) | Only compliant deployments | Requires legal and compliance sign-off |
⚠️ The Practical Rule
Before pasting anything into an AI system, ask: "Would I be comfortable if this appeared in a competitor's system?" If no — anonymise it or don't use it.
3. Human-in-the-Loop Systems
Human Review Triggers
- High-stakes decisions — affecting people's lives, safety, or significant financial outcomes
- Regulatory submissions — all AI-assisted content requires qualified specialist review
- Novel situations — when AI operates outside its verified use case
- Confidence flags — any output where the AI flagged uncertainty
- First-time outputs — before a new prompt type is trusted at scale
💡 The Verification Principle
For any AI output that will be acted on, a human must be able to trace the claim back to a verifiable source. If the output can't be traced, it can't be trusted — regardless of how confident the model sounded.
4. Prompt Governance in Organisations
Minimum Viable Governance Framework
- Approved use cases — define where AI can and cannot be used without review
- Data handling policy — what data can be entered into AI systems and under what conditions
- Output review requirements — which output categories require human expert review before use
- Prompt library — vetted, versioned prompts for common use cases, available to all staff
- Incident reporting — a mechanism to flag and learn from AI errors or near-misses
- Training — minimum prompt engineering literacy for all AI users in the organisation
✏️ Module 11 Exercise
Map your current AI use in your professional role. Apply the data classification framework to each use case. Design a minimum human-in-the-loop checkpoint for the highest-risk use case. Draft a one-page AI use policy for your team.
🔑 Key Takeaways
- AI bias is invisible and systematic — actively prompt for counterpoints and assumption checking
- Data sensitivity classification must precede AI use, not follow it
- Human-in-the-loop is not optional for high-stakes outputs — it is the design
- Prompt governance converts individual practice into organisational resilience
- The verification principle: if you can't trace it, you can't trust it
- Ethics in AI is about what humans decide to do with AI outputs — not just what AI can do