1 | 1 Decide who is accountable | 0 |
1.1 | 1.1 Accountable people | 0 |
1.1.1 | 1.1.1 Assign and communicate accountabilities for the AI management system | 2 |
1.1.2 | 1.1.2 Define competencies, authority and resources of accountable people | 0 |
1.2 | 1.2 Supply chain accountabilities | 0 |
1.2.1 | 1.2.1 Allocate shared responsibility across the AI supply chain | 0 |
1.2.2 | 1.2.2 Document developers' obligations to downstream organisations | 0 |
1.3 | 1.3 AI literacy and training | 0 |
1.3.1 | 1.3.1 Build organisation-wide AI literacy | 2 |
1.3.2 | 1.3.2 Train accountable people for their specific responsibilities | 0 |
1.4 | 1.4 AI governance framework | 0 |
1.4.1 | 1.4.1 Document strategic intent, applicable regulation and AI policies | 2 |
1.4.2 | 1.4.2 Operate the AI management system effectively | 1 |
1.4.3 | 1.4.3 Monitor compliance with AI policies | 1 |
2 | 2 Understand impacts and plan accordingly | 0 |
2.1 | 2.1 Identify and engage stakeholders | 0 |
2.1.1 | 2.1.1 Identify stakeholder types and their needs | 0 |
2.1.2 | 2.1.2 Select the stakeholder needs policies will address | 0 |
2.1.3 | 2.1.3 Commit to preventing harm and upholding diversity, inclusion and fairness | 1 |
2.1.4 | 2.1.4 Document each AI system's scope | 1 |
2.1.5 | 2.1.5 Engage stakeholders to identify benefits and harms | 1 |
2.1.6 | 2.1.6 Analyse the impact of every documented harm | 1 |
2.1.7 | 2.1.7 Monitor for harms through continuing stakeholder engagement | 0 |
2.1.8 | 2.1.8 Support meaningful stakeholder feedback | 1 |
2.2 | 2.2 Establish feedback and redress processes | 0 |
2.2.1 | 2.2.1 Establish a process for concerns, challenges and remediation | 1 |
2.2.2 | 2.2.2 Build system-level contestability mechanisms | 2 |
2.2.3 | 2.2.3 Let deployers escalate feedback to developers | 0 |
2.2.4 | 2.2.4 Enable reviewers to trigger recourse and redress | 0 |
2.3 | 2.3 Monitor for systemic issues | 0 |
2.3.1 | 2.3.1 Monitor contestability and redress for systemic risk | 0 |
2.3.2 | 2.3.2 Review stakeholder contests across the organisation | 0 |
3 | 3 Measure and manage risks: implement AI-specific risk management | 0 |
3.1 | 3.1 Establish a fit-for-purpose risk management framework | 0 |
3.1.1 | 3.1.1 Create an AI risk management framework, risk criteria and assessment processes | 1 |
3.1.2 | 3.1.2 Manage risks from third parties and open-source AI | 1 |
3.1.3 | 3.1.3 Report risk outcomes in consistent formats | 0 |
3.2 | 3.2 Assess AI system risks | 0 |
3.2.1 | 3.2.1 Triage AI systems for enhanced or unacceptable risk | 1 |
3.2.2 | 3.2.2 Assess each AI system's risks | 1 |
3.2.3 | 3.2.3 Evaluate likelihood and consequence, including of not deploying | 1 |
3.3 | 3.3 Implement controls for AI system risks | 0 |
3.3.1 | 3.3.1 Plan and implement risk treatment and verify it | 0 |
3.3.2 | 3.3.2 Communicate risk treatment plans | 0 |
3.3.3 | 3.3.3 Plan response and recovery for residual risk | 1 |
3.3.4 | 3.3.4 Apply leading-practice safety measures | 0 |
3.4 | 3.4 Monitor and report incidents | 0 |
3.4.1 | 3.4.1 Track and report serious incidents | 1 |
3.4.2 | 3.4.2 Meet AI-related reporting and disclosure obligations | 0 |
3.4.3 | 3.4.3 Conform to data breach reporting requirements | 1 |
3.4.4 | 3.4.4 Keep two-way incident communication between developers and deployers | 0 |
3.4.5 | 3.4.5 Review risk assessments and treatment plans | 1 |
3.4.6 | 3.4.6 Improve the risk management process | 0 |
4 | 4 Share essential information | 0 |
4.1 | 4.1 Maintain an AI register | 0 |
4.1.1 | 4.1.1 Maintain an organisation-wide AI register | 2 |
4.1.2 | 4.1.2 Adopt a transparency policy for AI use | 2 |
4.1.3 | 4.1.3 Set transparency and explainability requirements per system | 1 |
4.1.4 | 4.1.4 Scale transparency to general-purpose AI risks | 0 |
4.1.5 | 4.1.5 Prefer interpretable and explainable systems | 1 |
4.1.6 | 4.1.6 Provide explainable systems to accountable people and deployers | 0 |
4.1.7 | 4.1.7 Test and disclose capabilities and limitations before deployment | 0 |
4.2 | 4.2 AI system transparency and explainability | 0 |
4.3 | 4.3 Supply chain transparency | 0 |
4.3.1 | 4.3.1 Obtain technical details from upstream providers | 0 |
4.3.2 | 4.3.2 Share model and system information with downstream deployers | 0 |
4.3.3 | 4.3.3 Report issues and bias back to upstream developers | 0 |
4.3.4 | 4.3.4 Put information requirements in supplier contracts | 1 |
4.4 | 4.4 AI-generated content transparency | 0 |
4.4.1 | 4.4.1 Apply transparency mechanisms to AI-generated content | 1 |
5 | 5 Test and monitor | 0 |
5.1 | 5.1 Pre-deployment testing | 0 |
5.1.1 | 5.1.1 Oversee testing and post-deployment monitoring | 0 |
5.1.2 | 5.1.2 Define acceptance criteria and test before deployment | 1 |
5.1.3 | 5.1.3 Document tests and outcomes for audit | 1 |
5.1.4 | 5.1.4 Obtain and check supplier test results | 0 |
5.1.5 | 5.1.5 Obtain documented deployment authorisation | 0 |
5.2 | 5.2 Monitor system performance | 0 |
5.2.1 | 5.2.1 Monitor key metrics for each AI system | 2 |
5.2.2 | 5.2.2 Map business targets to performance metrics | 0 |
5.2.3 | 5.2.3 Establish response processes for foreseeable issues | 1 |
5.2.4 | 5.2.4 Hold regular performance reviews | 1 |
5.2.5 | 5.2.5 Set monitoring and oversight requirements before deployment | 0 |
5.3 | 5.3 Conduct additional testing proportionate to risk | 0 |
5.3.1 | 5.3.1 Conduct capability-scaled safety evaluations | 0 |
5.3.2 | 5.3.2 Independently review testing for enhanced-risk and GPAI uses | 1 |
5.3.3 | 5.3.3 Decide on and conduct risk-based audits | 1 |
5.4 | 5.4 Implement robust data and cybersecurity measures | 0 |
5.4.1 | 5.4.1 Extend data governance, privacy and cybersecurity policies to AI | 1 |
5.4.2 | 5.4.2 Apply data and security controls per AI use case | 1 |
6 | 6 Maintain human control | 0 |
6.1 | 6.1 Maintain human oversight and control | 0 |
6.1.1 | 6.1.1 Maintain accountability, capability and oversight across the lifecycle | 2 |
6.1.2 | 6.1.2 Build mechanisms for human control and intervention | 0 |
6.1.3 | 6.1.3 Provide oversight mechanisms for systemic risks and emerging capabilities | 0 |
6.1.4 | 6.1.4 Train people who oversee or use AI systems | 1 |
6.2 | 6.2 Decommission when appropriate | 0 |
6.2.1 | 6.2.1 Define termination criteria and plan decommissioning | 1 |
6.2.2 | 6.2.2 Notify affected parties of retirement or shutdown | 0 |
6.2.3 | 6.2.3 Provide alternative systems or processes where needed | 1 |
6.2.4 | 6.2.4 Maintain alternative pathways for critical functions | 2 |
X.1 | Nature, audience and versions | 0 |
X.2 | Supersession of the Voluntary AI Safety Standard and the AI Ethics Principles | 0 |
X.3 | Risk approach and AI-specific characteristics | 0 |
X.4 | Alignment with international standards | 0 |
X.5 | Official text | 0 |