Fairness
Assess whether outcomes differ unfairly across relevant groups. Document limitations and test performance with suitable data.
Practical controlDefine fairness measures, test by subgroup and record remediation decisions.
Trust by design
AI that earns trust, not just attention.
Responsible AI means building, deploying and monitoring AI systems with clear accountability, appropriate safeguards and people in the loop.
GovernanceRisk managementHuman oversightLifecycle controls
Core principles
Turn good intentions into practical controls.
Principles matter when they are translated into decisions, evidence, ownership and continuous monitoring.
Transparency
Make the system's purpose, intended use, limitations and role in decisions understandable to affected people.
Practical controlMaintain model/system documentation and communicate when AI is being used.
Privacy and security
Protect personal and confidential information and consider misuse, prompt injection, data leakage and supply-chain risks.
Practical controlApply data minimisation, access controls, security testing and incident response.
Accountability
Assign clear ownership for approving, operating, monitoring and retiring an AI system.
Practical controlDefine accountable owners, approval gates, audit evidence and escalation routes.
Safety and reliability
Evaluate performance under expected conditions and foreseeable failure or misuse scenarios.
Practical controlSet acceptance thresholds, red-team high-risk use cases and monitor drift and incidents.
Human oversight
Give people meaningful authority and enough information to review, challenge or override important outputs.
Practical controlDesign review paths, appeal mechanisms and fallback procedures for consequential decisions.
A lifecycle view
Governance is a continuous practice.
01
Discover and classify
Define purpose, users, impact, risk level and applicable requirements.
02
Design and build
Establish data controls, documentation, testing criteria and human oversight.
03
Validate and approve
Test accuracy, robustness, security, fairness and readiness before release.
04
Deploy and monitor
Track performance, drift, complaints, incidents and changes over time.
05
Improve or retire
Address findings, reassess risk and safely decommission systems when needed.
Useful frameworks
Structure for responsible practice.
NIST AI Risk Management Framework
A voluntary framework organised around Govern, Map, Measure and Manage to help organisations address AI risks.
ISO/IEC 42001
A management-system standard for establishing, implementing, maintaining and continually improving an AI management system.
OWASP Top 10 for LLM Applications
A practical reference for common security risks in applications that use large language models.
EU AI Act
A regulatory framework whose obligations depend on the system, role, risk category and applicable dates.
Frameworks and laws evolve. Confirm the latest text, applicability and effective dates for each specific use case. This page is educational information, not legal advice.
From principles to practice