목차
Title page
Contents
Foreword 2
Acknowledgements 3
Abstract 6
Resume 7
Background and objectives 8
Executive summary 9
Synthese 11
1. Introduction 13
1.1. The need for trustworthy AI 13
1.2. What is trustworthy AI? 13
1.3. What is accountability in AI? 14
2. DEFINE: Scope, context, actors, and criteria 18
2.1. Scope 18
2.2. Context 18
2.3. Actors 19
2.4. Criteria 22
3. ASSESS: Identify and measure AI risks 23
3.1. Benefiting people and the planet 23
3.2. Human-centred values and fairness 24
3.3. Transparency and explainability 29
3.4. Robustness, security, and safety 30
3.5. Interactions and trade-offs between the values-based Principles 31
4. TREAT: Prevent, mitigate, or cease AI risks 33
4.1. Risks to people and the planet 33
4.2. Risks to human-centred values and fairness 34
4.3. Risks to transparency and explainability 37
4.4. Risks to robustness, security, and safety 38
4.5. Anticipating unknown risks and contingency plans 40
5. GOVERN: Monitor, document, communicate, consult and embed 41
5.1. Monitor, document, communicate and consult 41
5.2. Embed a culture of risk management 49
6. Next steps and discussion 50
Annex A. Presentations relevant to accountability in AI from the OECD.AI network of experts 51
Annex B. Participation in the OECD.AI Expert Group on Classification and Risk 53
Annex C. Participation in the OECD.AI Expert Group on Tools and Accountability 55
References 58
Table 2.1. Sample processes and technical attributes per OECD AI Principle 22
Table 3.1. Examples of documentation to assess transparency and traceability at each phase of the AI system lifecycle 30
Table 4.1. Approaches to treating risks to people and the planet 33
Table 4.2. Approaches to treating bias and discrimination 34
Table 4.3. Approaches to treating risks to privacy and data governance 36
Table 4.4. Approaches to treating risks to human rights and democratic values 37
Table 4.5. Approaches to treating risks to transparency and explainability 37
Table 4.6. Approaches to treating risks to robustness, security, and safety 39
Table 5.1. Characteristics of AI auditing and review access levels 44
Figure 1.1. High-level AI risk-management interoperability framework 16
Figure 1.2. Sample uses of the high-level AI risk management interoperability framework 17
Figure 2.1. Actors in an AI accountability ecosystem 20
Figure 3.1. UK Information Commissioner's Office (ICO) qualitative rating for data protection 27
Figure 3.2. Mapping of algorithms by explainability and performance 32
Figure 5.1. Trade-off between information concealed and auditing detail by access level 45
Boxes
Box 1.1. What is AI? 13
Box 1.2. Trustworthy AI per the OECD AI Principles 14
Box 2.1. Mapping the lifecycle phases to the dimensions of an AI system 18
Box 3.1. Errors, biases, and noise: a technical note 25
Box 3.2. Human rights and AI 28
Box 3.3. Explainability vs interpretability 29
Annex Tables
Table A.1. OECD.AI expert presentations 51
Table B.1. Participation in the OECD.AI Expert Group on Classification & Risk (as of December 2022) 53
Table C.1. Participation in the OECD.AI Expert Group on Tools & Accountability (as of December 2022) 55
