Chapter 9. Responsible AI: Ethics, Fairness, Privacy, and Human Rights

9.1 What responsible AI means

Responsible AI means designing, procuring, and using systems in ways that respect people, reduce foreseeable harm, and make responsibility traceable. Common principles include validity, reliability, safety, security, transparency, explainability where appropriate, privacy, fairness, accountability, and human oversight. The OECD AI Principles emphasise human rights, democratic values, transparency, robustness, and accountability (OECD, n.d.-b).

Ethical principles must be translated into concrete practices: impact assessments, testing, access controls, documentation, human review, complaint routes, and monitoring. A statement of values is not enough if the operating process does not implement it.

9.2 Bias and fairness

Bias can enter through historical data, measurement choices, labels, sampling, model design, deployment context, or the way people respond to a system. A model trained on past decisions may reproduce unfair patterns even when protected characteristics are removed, because other variables can act as proxies.

Fairness is context-dependent. Different measures can conflict, and equal numerical outcomes do not always guarantee fair treatment. Teams should identify affected groups, consult relevant stakeholders, test subgroup performance where lawful and appropriate, examine error consequences, and document how trade-offs are resolved.

9.3 Privacy and confidentiality

Personal data should be collected and used only for legitimate, defined purposes, with appropriate safeguards and legal basis. Avoid entering names, account details, student records, medical information, private correspondence, passwords, unpublished research, or confidential business information into public AI tools unless explicitly authorised and protected by an approved arrangement.

Privacy protection includes more than removing names. Rare events, job titles, locations, and combinations of attributes may make a person identifiable. Use data minimisation, approved environments, access controls, retention limits, and secure deletion. Check the applicable law and institutional requirements.

9.4 Transparency and explainability

People should receive meaningful information when AI materially affects them, appropriate to the context and law. Explainability may involve describing the purpose, inputs, main factors, limitations, human role, and how to challenge a result. A technical explanation that ordinary users cannot understand may not provide meaningful transparency.

Some generative AI outputs are difficult to explain at the level of individual tokens. Even so, organisations can document the model and version, source materials, prompt or workflow, test results, known limitations, approval process, and responsible owner.

9.5 Human oversight and accountability

Human oversight must be real rather than ceremonial. Reviewers need enough time, competence, information, authority, and independence to question an AI result. If staff are expected to approve every output at high speed without the ability to inspect evidence, the review may become a rubber stamp.

Responsibility should be assigned before deployment. Name the business owner, technical owner, data owner, risk contact, and final decision-maker. Establish escalation routes and define which actions require explicit approval. The organisation remains responsible for its use of the system even when a vendor provides the model.

9.6 Inclusion and accessibility

Systems may perform differently across languages, dialects, disability contexts, ages, regions, and levels of digital access. Test with the populations expected to use or be affected by the tool. Provide non-AI alternatives where necessary and ensure that people can access services without being unfairly excluded because they cannot use a particular interface.