Welcome to the final track. Everything you have learned — from gradient descent to transformers — is powerful precisely because it scales: one model can make millions of decisions. That power is why ethics is not an optional appendix to AI education. A biased model does not discriminate once; it discriminates systematically, at scale, often invisibly. As engineers, we are responsible not only for whether our systems work, but for whom they work, how, and with what consequences.
Real harms, not hypotheticals#
Documented cases include:
- Criminal justice: ProPublica's 2016 analysis of the COMPAS recidivism tool reported that Black defendants who did not reoffend were almost twice as likely as white defendants to be labelled high risk — igniting a debate about fairness definitions that continues today.
- Hiring: Amazon reportedly abandoned an experimental résumé-screening tool after finding it penalised résumés containing the word "women's", having learned from historically male-dominated hiring data.
- Healthcare: Obermeyer et al. (2019, Science) found that a widely used algorithm for allocating extra care used past health costs as a proxy for health needs; because less money had historically been spent on Black patients with the same needs, the algorithm substantially under-identified Black patients who needed extra care.
- Face recognition: Buolamwini and Gebru's Gender Shades (2018) found commercial gender classifiers had far higher error rates for darker-skinned women; wrongful arrests based on facial-recognition matches have been reported.
- Public services: the Dutch childcare-benefits scandal involved risk-scoring and profiling practices that wrongly accused thousands of families of fraud, with devastating consequences, contributing to the government's resignation in 2021.
In each case, the technology "worked" by some metric — and still caused serious harm.
Where harms come from#
Harms can enter at every stage of the ML lifecycle (Suresh & Guttag, 2021):
- Problem formulation: choosing a proxy target (cost instead of need; arrests instead of crime).
- Historical bias: data reflecting past discrimination.
- Representation bias: some groups under-represented in data.
- Measurement bias: features or labels measured differently across groups.
- Aggregation bias: one model for groups with different relationships.
- Evaluation bias: benchmarks not representative of the deployment population.
- Deployment bias: using a system in contexts it was not designed for, or in ways that shift power.
Principles of responsible AI#
Hundreds of AI ethics guidelines have been published; analyses (e.g. Jobin, Ienca & Vayena, 2019) found convergence around a few principles:
| Principle | Meaning in practice |
|---|---|
| Fairness / non-discrimination | Equitable performance and outcomes across groups |
| Transparency / explainability | People can understand how systems work and why decisions were made |
| Accountability | Clear responsibility; ability to contest and seek redress |
| Privacy | Data minimisation, consent, security, purpose limitation |
| Safety / robustness | Reliable behaviour, resistance to misuse and attack |
| Beneficence / human wellbeing | Systems should genuinely benefit people |
| Human autonomy / oversight | Humans remain in meaningful control of consequential decisions |
Humanitarian and development work adds principles such as "do no harm", humanity, impartiality, and particular care for the dignity and protection of people in vulnerable situations.
From principles to practice#
Principles are easy to endorse and hard to implement. Concrete practices include:
- Stakeholder engagement: involve affected communities and domain experts from problem framing onward.
- Impact assessments before building and before deployment.
- Data documentation (datasheets) and model documentation (model cards) with disaggregated evaluation.
- Fairness testing across relevant groups.
- Explanations and contestability — ways for people to question and appeal decisions.
- Human oversight with real authority and time to exercise judgement (not rubber-stamping).
- Monitoring after deployment, including for harms.
- The option not to build: sometimes the responsible choice is that a problem should not be automated.
Ethics is part of engineering quality#
A model that is 95% accurate overall but 70% accurate for a minority language, or that can be easily manipulated, or that leaks personal data, is a defective product. Responsible AI is not in tension with good engineering; it is a dimension of it. The rest of this track gives you tools: fairness metrics, explainability, privacy techniques, safety, regulation, and career guidance for building AI that serves people well.