Singapore Approach To Developing Trustworthy AI

Singapore Approach To Developing Trustworthy AI Presented by: Yeong Zee Kin, Deputy Commissioner, Personal Data Protection Commission, Singapore Date: 22 November 2021 Singapore's National Al Strategy - VISION • By 2030, Singapore will be a leader in developing and deploying scalable, impactful Al solutions, in key sectors of high value and relevance to our citizens and businesses. - - NATIONAL AI PROJECTS • Transport and logistics • Smart cities & estates • Healthcare • Education • Education • Safety & security ECOSYSTEM ENABLERS • Triple helix partnership • AI talent & education • Data architecture • Progressive & trusted environment • International collaboration Building the Al Ecosystem - Even as Singapore embarks on National Al Projects, we must build a vibrant and sustainable Al ecosystem to anchor Al innovation and adoption across the economy. - We have identified five ecosystem enablers: • Triple helix partnership • AI talent and education • Data architecture • Progressive and trusted environment • International collaboration Singapore's Approach to Al Governance - Industry voluntary adoption of responsible Al with detailed government guidance - Government guidance - • Model Al Governance Framework (2nd edition) • Implementation & Self-Assessment Guide for Organizations • 2 volumes of Compendium of Use Cases Multi-stakeholder approach • Living documents that will evolve with tech development • Input and feedback from >60 companies of different sizes, from different sectors, locally and internationally • Worked with int'l organizations like WEF, OECD Model Al Framework is a Baseline - Accountability-based framework to engender trust: Helps companies be transparent about their governance practices in developing & deploying Al systems - Takes Al ethics into corporate governance, risk management and operational structures - Baseline for other sectors to build on - Key characteristics • Sector-agnostic • Technology-agnostic • Algorithm-agnostic • Scale- and business-model-agnostic • From principles to practice: practical, implementable • Does not define ethical standards Accountability-Based Practices & Measures - - - - - Guiding Principles • Decision making process should be explainable, transparent and fair • Al systems should be human-centric and fair INTERNAL GOVERNANCE STRUCTURES AND MEASURES • Corporate governance structures • Oversight mechanisms • Roles and responsibilities • Monitoring and reporting systems and processes • Periodic review HUMAN INVOLVEMENT IN AI-AUGMENTED DECISION-MAKING • Risk and impact assessments • incorporating ethical considerations • Degree of human intervention in Al • decision-making • Risk mitigating measures OPERATIONS MANAGEMENT • Data curation, model formulation, model selection, training and tuning • Ensure quality, credibility of source, veracity, while mitigating inherent bias • Explainability, repeatability & robustness • Reproducibility, auditability for certain scenarios STAKEHOLDER INTERACTION & COMMUNICATION • Build and maintain open relationship with stakeholders that builds trust • Communication appropriate to audience, purpose and context • Communication channels NEW: Al Governance Testing Framework - Demonstrate responsible Al through testing Al system against ethical principles objective - Help companies build trust with their stakeholders and meet regulatory requirements - UNDERSTANDING HOW AI MODEL REACHES DECISION • To know what it does and that results are consistent • 1. EXPLAINABILITY  Ability to understand and interpret what the Al system is doing • 2. REPEATABILITY  Check that it's consistent: Be able to replicate an Al system's results • 3. REPRODUCIBILITY  Ability to replicate an Al system's results (by independent 3rd-party) - SAFETY AND RESILIENCE OF AI SYSTEM • Al system is reliable and will not cause harm • 4. SAFETY  Check that it's safe: Known risks have been identified/mitigated • 5. SECURITY  Cybersecurity of Al systems • 6. ROBUSTNESS  Ensuring that Al system can still function despite unexpected inputs - FAIRNESS / NO UNINTENDED DISCRIMINATION • Al system does not unintentionally discriminate • 7. FAIRNESS  Check that there is no unintended bias: Al systems makes same decision even if an attribute is changed • 8. DATA GOVERNANCE  Know the source and quality of data: Good data governance practices when training Al models - MANAGEMENT AND OVERSIGHT OF AI SYSTEM • Human accountability and control • 9. ACCOUNTABILITY  Proper management oversight of Al system development • 10. TRANSPARENCY  Appropriate info is provided to individuals impacted by Al system • 11. HUMAN AGENCY AND OVERSIGHT  Al system designed in a way that will not decrease human ability to make decisions • 12. INCLUSIVE GROWTH, SOCIETAL & ENVIRONMENTAL WELLBEING  Beneficial outcomes for people and planet Data is the Building Block of Al - Trusted data flows are core to building an Al ecosystem - Personal Data Protection Act (PDPA) and obligations such as consent, accountability, access/correction, retention limitation apply to Al systems relying on personal data - Enhanced consent framework to enable organizations to harness data confidently to increase business competitiveness while strengthening consumer trust in Singapore's Digital Economy - Providing Greater Ease and Certainty to Use Data for Innovation: • Business Improvement Exception  (I) Operational efficiency & service improvements  (II) Developing or enhancing products or services  (III) Knowing the organizations’ customers • Legitimate Interests Exception  Lawful interests of organization or segment of public, such as anomaly detection in payment systems to prevent fraud or money-laundering Summary of Singapore's Approach to Al Governance - Partnership: Collaboration with key stakeholders in government, industry, consumers and academia, • - e.g. WEF and its network, tech MNCs Practicality: Operationalization of Al ethics principles through implementable practices • e.g. Organizations from various sectors, both local and overseas, putting in place the responsible Al practices outlined in the Framework - Participation: Strong support for open and interoperable international environment and regime, enabling human-centric Al innovation and its adoption • e.g. contribution to international platforms like GPAI, OECD