Singapore: Singapore Model AI Governance Framework, Second Edition (PDPC/IMDA)

1. ANNEX A
2. ANNEX B
3. 2019 Edition) at the 2019 World Economic Forum Annual
4. Annex A – For reference: a compilation of existing
5. Annex C – Use Case: • Annex C has been removed. Instead, a separate
6. 1. PREAMBLE
7. 1.2 It is recognised that there are a number of issues that are closely
8. 2. INTRODUCTION
9. 2.1 The exponential growth in data and computing power has fuelled
10. 2.2 The PDPC,1 with advice from the Advisory Council, proposes this
11. 2.3 This voluntary Model Framework provides guidance on the key
12. 2.5 The extent to which organisations adopt the recommendations
13. 2.6 To elaborate: AI technologies may be used to augment a human
14. 2.7 The Model Framework is based on two high-level guiding
15. 2.9 Organisations should detail a set of ethical principles when they
16. 2.10 The Model Framework aims to discuss good data management
17. 2.11 The Model Framework does not address the risk of catastrophic
18. 2.12 Adopting this voluntary Model Framework will not absolve
19. 2.13 Further, it should be noted that certain industry sectors (such
20. 2.14 The following simplified diagram depicts the key stakeholders
21. 2.15 Some terms used in AI may have different definitions depending
22. 3. MODEL AI
23. 3.1 This Model Framework comprises guidance on measures
24. 3.2 Organisations adopting this Model Framework may find that not
25. 3.4 This section is intended to guide organisations in developing
26. 3.5 Internal governance structures and measures help to ensure robust
27. 3.6 Organisations may also consider determining the appropriate
28. 3.7 Organisations may wish to consider including features that
29. 1. Clear roles and responsibilities for the ethical
30. 2. Risk management and internal controls
31. 3.8 This section is intended to help organisations determine the
32. 3.9 Having clarity on the objective of using AI is a key first step in
33. 3.10 It is also desirable for organisations operating in multiple countries
34. 3.11 Some risks to individuals may only manifest at group level. For
35. 3.12 Organisations’ weighing of their commercial objectives against
36. 3.13 As identifying commercial objectives, risks and determining the
37. 3.14 Based on the risk management approach described above, the Model Framework
38. 3.15 The Model Framework also proposes a design framework
39. 3.16 The definition of “harm” and the computation of probability and
40. 3.17 The matrix, however, should not be taken to imply that the
41. 3.18 For safety-critical systems, it would be prudent for organisations
42. 3.19 This section is intended to help organisations adopt responsible
43. 3.20 The Model Framework uses the following generalised AI model
44. 3.21 During deployment, algorithms such as linear regression
45. 3.22 Datasets used for building models may come from multiple
46. 3.23 The persons who are involved in training and selecting models
47. 3.24 Even if only non-personal data are used for the training of AI
48. 3.25 AI systems may have numerous features or functionalities enabled
49. 3.26 Explainability is achieved by explaining how deployed AI models’
50. 3.27 Organisations deploying AI solutions are recommended to adopt
51. 3.28 Technical explainability may not always be enlightening, especially
52. 3.29 Nevertheless, there may be scenarios where it might not be
53. 6 These tools are known as “supplementary” as there is at present no single comprehensive
54. 3.30 Where explainability cannot practicably be achieved given the
55. 7 James Manyika, Jake Sitberg, and Brittany Presten, “What Do We Do About the Biases in
56. 3.31 Robustness refers to the ability of a computer system to cope
57. 3.32 The concept of robustness arises because it is not possible for
58. 3.33 No model can be perfectly robust as it is not possible to detect
59. 8 This is distinct from user acceptance testing (“UAT”), which is a process where actual
60. 3.34 Establishing an internal policy and process to perform regular
61. 3.35 Wherever possible, testing should reflect the dynamism of the
62. 3.36 An AI model is considered to be traceable if (a) its decisions, and
63. 3.37 Practices that organisations may consider to promote traceability
64. 3.38 As traceability measures may lead to an accumulation of a large
65. 9 It should be noted that a black box recorder does not refer to a “black box” in the AI
66. 3.39 While repeatability refers to the internal repetition of results within
67. 3.41 Auditability refers to the readiness of an AI system to undergo
68. 3.42 Implementing auditability not only entails the involvement of
69. 3.43 To facilitate auditability, organisations can consider keeping a
70. 3.44 Organisations may also wish to centralise such information digitally
71. 3.45 This section is intended to help organisations take appropriate
72. 3.46 Organisations are encouraged to provide general information
73. 3.47 Organisations can consider disclosing the manner in which an
74. 3.48 Organisations are encouraged to develop a policy on what
75. 3.49 Appropriate interaction and communication inspire trust and
76. 3.50 As different stakeholders have different information needs, an
77. 3.51 Organisations are encouraged to consider the information needs
78. 3.52 Organisations may wish to consider carefully when deciding
79. 3.53 Where an organisation has weighed the factors above and decided
80. 3.54 Organisations are encouraged to put in place the following
81. 3.55 Organisations are encouraged to test user interfaces and address
82. 3.56 Organisations are encouraged to communicate in an easy-to-
83. 3.57 In certain cases, organisations may be implementing AI-powered
84. 3.58 In this regard, AUPs serve to set broad boundaries for the
85. 3.59 Some of the approaches and methodologies described in the
86. 10 For example, information related to lineage of the training dataset and documenting the
87. 3.60 Organisations may have to consider the level of support and
88. 3.61 Depending on the purpose and context, the type and level of
89. 3.62 Finally, as ethical standards governing the development and use
90. 3.63 This Model Framework is by no means complete or exhaustive
91. 1. Accountability: Ensure that AI actors are responsible and accountable for the proper
92. 2. Accuracy: Identify, log, and articulate sources of error and uncertainty throughout
93. 3. Auditability: Enable interested third parties to probe, understand, and review the
94. 4. Explainability: Ensure that automated and algorithmic decisions and any associated
95. 7. Human rights alignment: Ensure that the design, development and implementation of
96. 9. Progressiveness: Favour implementations where the value created is materially better
97. 11. Robustness and Security: AI systems should be safe and secure, not vulnerable to
98. 12. Sustainability: Favour implementations that effectively predict future behaviour and
99. 1. Algorithm audits are conducted if it is necessary to discover the actual operations of

The outline comes from headings detected in the text, so it might miss some or include extras.