AI ethics

In plain English

Law tells you what you must do. Ethics asks what you should do. A building code sets the minimum for safety. A good architect also thinks about light, access and the neighbors.

AI ethics asks questions like these. Who could this system harm? Is it fair to every group it touches? Can people understand and challenge its decisions? Who answers when it goes wrong?

Principles on a poster change nothing. Our lesson I.A.5 traces the chain that makes them real: principle, policy, procedure, control, evidence.

Example

Larchmont Mutual, a fictional insurer, plans to price home insurance with a model using satellite images and customer purchase data. Its ethics review starts at the planning stage, not the week before launch.

• Fairness. Premiums are tested across neighborhoods for proxies of race.
• Privacy. The team asks whether purchase data is needed at all.
• Transparency. Customers will be told AI is used and which data feeds it.
• Accountability. A named executive owns the model and can suspend it.

Then two values collide. The most accurate version relies on purchase data that tracks a protected trait. Larchmont quantifies both options and checks anti-discrimination law as the floor. A cross-functional group decides, and the rationale is documented. It drops the purchase data and monitors outcomes after launch.

Why it matters

The major frameworks agree on the core values:

The EU expert group lists seven requirements:

1. Human agency and oversight
2. Technical robustness and safety
3. Privacy and data governance
4. Transparency
5. Diversity, non-discrimination and fairness
6. Societal and environmental well-being
7. Accountability

Recital 27 of the EU AI Act recalls them as non-binding. The Act's rules for high-risk systems give several of them legal force, such as human oversight (Art. 14) and data governance (Art. 10).

How the AIGP exam tests it

Our lessons group the BoK's principles into six pillars of responsible AI (lesson I.A.4). They are fairness, safety and reliability, privacy and security, transparency and explainability, accountability, and human-centricity. It then tests what happens when they collide (lesson I.A.5).

Common traps:

• Profit as a principle. No framework lists profit, speed or market share.
• Contracting out accountability. An indemnity moves money, not accountability.
• Pick a winner. The right answer balances values, decides cross-functionally and documents why.
• Leave it to the data scientists. It is an ethical and legal decision in statistical clothing.
• Ethics at the end. A pre-launch ethics gate is the near miss; ethics by design starts at planning.

The tell: legality is the floor, not the ceiling.