Building an AI system that works is half the job; building one that is fair, safe and accountable is the other half. This section covers the principles, lifecycle controls and tools of responsible AI.
Responsible AI is the set of practices that keep machine learning and generative AI systems fair, safe, transparent, private and accountable throughout their lives. It is not a separate phase bolted on at the end; it is a set of questions and controls attached to every stage of building a system, from deciding whether to build it at all to monitoring it years after launch.
Most organisational frameworks, including the NIST AI Risk Management Framework and the OECD AI Principles, converge on roughly the six principles above. Regulation is also arriving: the EU AI Act classifies systems by risk and imposes obligations such as documentation, human oversight and accuracy requirements on high-risk uses like hiring and credit scoring.
The four pages in this section map onto this picture: fairness and explainability are mostly build-time concerns, safety and alignment dominate evaluation and deployment of generative models, and governance spans the whole loop. The lifecycle view is the practical one. At definition, teams state the intended use, who could be harmed, and the risk tier, which decides how much of the rest is mandatory. In data, they check consent, provenance and representation, because a model cannot be fairer than the data allows. During build, they measure performance per subgroup and choose explainable methods where decisions affect people. Evaluation adds adversarial testing: red teams try to make a model leak data, produce harmful content or be jailbroken. Deployment adds runtime guardrails, rate limits and human review for high-stakes outputs. Monitoring watches for drift, collects incident reports and feeds findings back into the next iteration.
For traditional ML the emphasis falls on fairness metrics, explainability (SHAP values, counterfactual explanations) and data governance. For large language models it shifts toward content safety, hallucination, prompt injection, privacy of training and retrieval data, and evaluation of open-ended outputs.
| Concern | Typical question | Main techniques | Applies most to |
|---|---|---|---|
| Fairness | Does accuracy differ across groups? | Subgroup metrics, reweighting, threshold adjustment | Credit, hiring, healthcare models |
| Transparency | Can we explain a decision? | Model cards, SHAP, interpretable models | Regulated decisions |
| Privacy | Could the model reveal personal data? | Data minimisation, differential privacy, PII redaction | Models trained on user data, RAG |
| Safety | Could outputs cause harm? | Red teaming, content filters, refusal training | Generative AI, agents |
| Robustness | Does it fail under attack or shift? | Adversarial testing, drift monitoring | Security-sensitive systems |
| Accountability | Who answers when it goes wrong? | Named owners, audit logs, incident process | Every production system |
A useful rule of thumb: the more a system's output directly changes something about a person's life without human review, the higher its risk tier and the more of this table becomes mandatory rather than optional.