1-4 | Scope, terms and abbreviations | 0 |
5.1-5.14 | AI concepts: agents, knowledge, learning paradigms, autonomy and cyber-physical systems | 0 |
5.10 | Run the data processes an AI system depends on | 0 |
5.11.6-5.11.8 | Keep training, validation and test data separate and representative | 0 |
5.11.9 | Retrain for data drift and concept drift, and govern continuous learning | 0 |
5.15.1 | What trustworthiness lets stakeholders verify | 0 |
5.15.2 | Establish AI robustness: performance held under atypical conditions | 0 |
5.15.3 | Establish AI reliability over the operation stage | 0 |
5.15.4 | Establish AI resilience and fault tolerance with recovery | 0 |
5.15.5 | Provide controllability: who can intervene in which component | 0 |
5.15.6 | Provide explainability of the factors behind decisions | 0 |
5.15.7 | Define predictability with objective criteria | 0 |
5.15.8 | Provide transparency to stakeholders about the system and its data | 0 |
5.15.9 | Detect and treat unwanted bias to avoid unfairness | 0 |
5.16 | Decide what can be verified and what must be validated | 0 |
5.17 | Plan for legal requirements that differ between jurisdictions | 0 |
5.18 | Assess the level of societal risk of the system | 0 |
5.19 | Identify the AI stakeholder roles the organisation and others hold | 0 |
6.1 | Adopt a life cycle model that carries governance, privacy, security and transparency throughout | 0 |
6.2.2a | Inception: set objectives and life-cycle-wide requirements | 0 |
6.2.2b | Inception: assess AI risk, identify harms, benefits and values, and plan treatment | 0 |
6.2.2c | Inception: record for transparency and accountability, and secure cost, funding and resources | 0 |
6.2.3 | Design and development: approach, architecture, code, training data and risk treatment | 0 |
6.2.4 | Verification and validation: test with separate representative data, accept, review risk | 0 |
6.2.5 | Deployment: target environment and risk treatment review | 0 |
6.2.6 | Operation and monitoring: monitor, repair, update, support and review risk | 0 |
6.2.7 | Continuous validation where the system keeps learning | 0 |
6.2.8 | Re-evaluation: assess results against objectives and refine them | 0 |
6.2.9 | Retirement: decommission and discard, or replace | 0 |
7.1 | Provide human design choices, engineering and oversight throughout | 0 |
7.2-7.3 | Functional view: data and information, knowledge and learning | 0 |
7.4 | Understand that outputs are error prone and set thresholds for automated decisions | 0 |
8-10 | AI ecosystem layers, fields of AI and example applications | 0 |
A | Annex A: Mapping of the life cycle to the OECD definition | 0 |