Ethics in practice
The hard part of responsible AI is deciding what to do when principles pull in different directions. This module gives you a practical, repeatable approach.
Learning objectives
- Recognise common tensions between principles
- Apply a seven-step framework to ethical dilemmas
- Map responsible AI activities to each lifecycle stage
- Describe who is responsible for what
- Plan meaningful stakeholder engagement
6.1 Principles conflict
Real projects rarely let you maximise every principle at once. Common tensions:
| Tension | Example |
|---|---|
| Fairness vs privacy | Testing for bias needs demographic data you might otherwise not collect |
| Accuracy vs explainability | A complex model is slightly more accurate but much harder to explain |
| Transparency vs security | Publishing details of a fraud model helps fraudsters evade it |
| Helpfulness vs safety | An assistant that refuses too much is useless; one that refuses too little enables harm |
| Innovation vs caution | Waiting for certainty means forgoing real benefits |
6.2 A framework for ethical dilemmas
When principles conflict, work through these steps — ideally with a diverse group:
- Frame the decision. What exactly are we deciding? What's the purpose of the AI?
- Identify stakeholders. Who's affected — directly and indirectly? Who isn't in the room?
- Map benefits and harms for each group — their likelihood, severity, reversibility and distribution.
- Apply the principles. Which are in tension? What does each suggest?
- Generate options, including "don't use AI" and "use it with extra safeguards".
- Decide and justify. Choose, record why, and name who's accountable.
- Monitor and revisit. Set triggers that would make you reconsider.
Situation: A housing association wants AI to predict which tenants are at risk of falling into rent arrears, so it can offer early support.
Benefits: earlier help, fewer evictions. Harms: tenants feeling surveilled; predictions used punitively; bias against groups with less stable income; sensitive inferences.
Possible decision: proceed only for supportive outreach, never for enforcement; exclude sensitive data; test for disparities across groups; tell tenants about the scheme; let them opt out; review results with a tenant panel after six months.
6.3 Building responsible AI into the lifecycle
| Stage | Responsible AI activities |
|---|---|
| Idea | Is AI the right tool? Initial impact screening; stakeholder input |
| Design | Impact assessment; choose fairness metrics; plan human oversight; DPIA |
| Data | Representativeness, consent and provenance checks; datasheets |
| Build & test | Disaggregated evaluation; red-teaming; model cards |
| Deploy | Disclosure; user training; contestability routes; go/no-go review |
| Operate | Monitoring; incident response; periodic review; decommissioning |
6.4 Who does responsible AI?
Everyone has a part:
- Leaders set priorities, resources and culture.
- Product and engineering teams build responsible practices into their work.
- Legal, privacy, risk and security teams advise and provide assurance.
- Ethics boards or responsible AI councils review high-risk cases.
- Users and affected communities provide input and feedback.
- Every employee can raise concerns — and should be able to without fear.
6.5 Engaging stakeholders
The people most affected by AI are often least represented in building it. Techniques include user research with diverse participants, community panels, consultations, accessible feedback channels and involving workers in decisions about AI at work.
Choose an AI use case — real or imagined (e.g. AI to prioritise GP appointment requests; AI to screen supplier invoices for fraud; a customer chatbot for a bank). Work through the seven-step framework. Write a one-page decision record: stakeholders, benefits and harms, principles in tension, decision, safeguards, owner and review date.
Key takeaways
- Principles conflict — fairness vs privacy, accuracy vs explainability, helpfulness vs safety.
- Frame, identify stakeholders, map harms and benefits, apply principles, generate options, decide and justify, monitor.
- Build responsible AI into every lifecycle stage, from idea to decommissioning.
- Everyone has a role; people must be able to raise concerns safely.
- Involve those most affected.