Creating ethical AI in practice
The operational checklist for deciding which use cases meet an organisation's ethical principles - spanning legal review, equitable design, transparency, privacy & cybersecurity, data governance and culture. Data minimisation and explainable third-party decisions are recurring exam points.
The operational checklist → a process to decide which use cases meet the organisation's ethical principles and code of conduct.
- Legal & compliance → guidance, policies and procedures ensuring legal review of AI and execution of bias-mitigation processes. If none exists, develop it.
- Equitable design → diversity of thought in the teams that develop, train, test and monitor AI → no diversity means biased inputs or outcomes are more likely. Higher-risk products need a cross-functional, demographically diverse review group.
- Transparency & interpretability → label AI systems internally and externally per FTC guidance; notify consumers when they interact with AI; decisions must be explainable to the consumer, including when AI comes from a third party → due diligence and contracts must secure those explanations; people may seek human intervention in decisions affecting legal rights or well-being.
- Privacy & cybersecurity → disclose AI training uses of personal data in privacy notices; obtain consent for automated profiling per GDPR, CCPA, US state laws, Brazil's LGPD; honour access and deletion rights; data minimisation → exclude personal data unlikely to improve the model; defend against extraction of personal data and poisoning of the model.
- Data governance → ensure the quality and integrity of the data used to develop and train models.
- Culture → programmes to train and educate employees → a culture of ethical AI, not just a policy binder.
Module 2's five takeaways: 1) know the OECD principles (five for trustworthy AI) plus the 1980 FIPs they grew out of; 2) recognise the five harm targets (individuals, groups, society, organisations, ecosystems) - harms can hit several at once; 3) use the harms taxonomies (PANOPTIC, Calo, Citron & Solove for privacy, Sociotechnical, CSET, NIST for AI); 4) address bias proactively (implicit, sampling, temporal, overfit, underfit, edge cases, legal vs illegal bias); 5) emphasise transparency (explainable systems, labelled AI, challengeable outcomes). Mnemonic bank → HAT, People Rarely Recall Every Detail, Psychology Aside, Inclusive Humans Trust Robust Accountability.
Key terms - quick answers
What is “Data minimisation”?
Sources and study method
This independent lesson uses active recall, spaced retrieval and scenario practice. Read the full study method.