AI Ethics

Quick reference notes for the COMP3007 AI ethics content.

1. Core AI Ethics Principles (Australian AI Ethics Principles)

We may be asked to identify which principle is demonstrated in a short scenario.

Key examples to remember:

• Human, societal and environmental well-being
  • Example: An AI system that helps monitor bushfires to reduce environmental damage.
  • Focus: long-term benefits, minimising harm to people, communities, and the environment.
• Fairness
  • Example: Consulting stakeholders when designing an AI system so diverse perspectives are included.
  • Focus: avoiding bias, ensuring equitable treatment and outcomes.
• Human-centred values
  • Example: CBA’s “Bill Sense” feature that:
    • Lets users control inputs,
    • Opt in or remove predictions.
  • Focus: autonomy, dignity, user control and empowerment.
• Privacy protection and security
  • Example: Flamingo AI:
    • Uses only necessary data,
    • Prevents unauthorised access or use.
  • Focus: data minimisation, security controls, safeguarding information.
• Contestability
  • Example: IAG allows customers to use internal and external complaint processes to challenge claims decisions.
  • Focus: users can challenge, appeal or dispute AI-influenced outcomes.

Other principles (awareness level for MCQs):

• Transparency & explainability – explaining how AI decisions are made.
• Reliability & safety – robust, tested, safe systems.
• Accountability – people and organisations remain responsible for impacts.

2. Voluntary AI Safety Standard – Guardrails

We’ll likely get one multiple-choice question about the 10 Guardrails.

Memorise these examples:

• Guardrail 1 – Accountability, governance & compliance
  • Example: A project with a dedicated team, clear timelines, and risk management expertise, plus a strategy for regulatory compliance.
  • Keywords: governance, accountability, internal capability, compliance plan.
• Guardrail 4 – Test & monitor AI models and systems
  • Example: An AI model whose performance is regularly monitored based on real-world user interactions and safety metrics.
  • Keywords: testing, evaluation, ongoing monitoring after deployment.
• Guardrail 7 – Contestability processes
  • Example: Creating and communicating a process for impacted people to:
    • Raise concerns,
    • Request remediation,
    • Contest AI decisions.
  • Keywords: challenge, appeal, recourse, complaints.

Remember:

• Data protection → Privacy & Security (not necessarily Guardrail 7).
• Complaint/appeal flows → Contestability (Guardrail 7 & contestability principle).

3. AI Impact Navigator – Plan–Act–Adapt & 4 Dimensions

3.1 Plan–Act–Adapt Cycle

The Plan–Act–Adapt cycle is a continuous improvement loop:

• Plan – identify potential AI impacts, set goals and safeguards.
• Act – deploy systems with controls and ethical practices in place.
• Adapt – monitor outcomes and update systems based on feedback and evolving risks.

MCQ summary:

“A continuous improvement cycle to guide measurement, action, and learning for AI impacts.”

3.2 The 4 Impact Dimensions

Know which dimension deals with transparency and public trust:

• Social licence & corporate transparency
  • Focus: building public trust and stakeholder confidence through:
    • Openness about AI use,
    • Clarity on data handling and impacts,
    • Transparent decision-making.

Other dimensions (awareness):

• Impacts on people and communities.
• Impacts on the environment.
• Governance, risk and compliance.

4. Quick Q&A Recap

A last-minute checklist:

1. Bushfire monitoring system → Human, societal and environmental well-being
2. Consulting stakeholders during design → Fairness
3. Bill Sense with opt-in/out & control over predictions → Human-centred values
4. Flamingo AI limiting data to what’s necessary & preventing misuse → Privacy protection and security
5. IAG customers can challenge claims decisions internally & externally → Contestability
6. Project with clear governance team & regulatory strategy → Guardrail 1 – Accountability, governance & compliance
7. Monitoring model performance with real-world data & safety metrics → Guardrail 4 – Test and monitor AI systems
8. Process for impacted people to raise concerns and contest outcomes → Guardrail 7 – Contestability
9. Plan–Act–Adapt → Continuous improvement cycle for measuring, acting and learning from AI impacts
10. Dimension focused on transparency and trust → Social licence & corporate transparency