Singapore AI Governance Framework

AI provides recommendations while a human makes the final decision and is accountable

• HITL procedure documentation
• decision review records
• training documentation
• override logs

• no HITL procedure
• weak training
• missing override logs

AI makes decisions but a human can override or intervene when necessary

• HOTL procedure documentation
• monitoring records
• intervention logs
• training documentation

• no HOTL design
• weak monitoring
• missing intervention logs

AI makes decisions autonomously; appropriate only when probability and severity of harm are low

• HOOTL risk assessments
• guardrails documentation
• monitoring procedures
• kill switch records

• no HOOTL risk assessment
• weak guardrails
• missing kill switch

Determine appropriate level of human involvement based on probability and severity of harm to individuals

• risk-severity methodology
• AI tier assignments
• control mapping per tier
• review records

• no methodology
• weak tier assignments
• missing review

Per MAF + GenAI Companion Paper: GenAI specifics. Requirements include (a) GenAI Output Verification and Safe Use + (b) Provenance and Watermarking of AI-Generated Content + (c) hallucination + misuse prevention + (d) maintain documentation.

• SG AI Gov evidence for SGAIGOV-5

• AI ethics body + GenAI provenance partial

Per Singapore Model AI Governance Framework (MAF) v2: governance. Requirements include (a) Internal Governance Structures and Measures + AI Ethics Governance Body + (b) Risk Management and Internal Controls + (c) Roles + Responsibilities + Accountability + (d) align with PDPC + IMDA guidance.

• SG AI Gov evidence for SGAIGOV-1

• AI ethics body + GenAI provenance partial

Establish clear internal governance structures for AI deployment including risk management and internal controls

• AI risk management procedures
• internal control documentation
• audit records
• executive reports

• no AI-specific controls
• weak audit
• missing executive reports

Set up an ethics review board or governance body for oversight of AI systems

• AI ethics committee charter
• membership records
• meeting minutes
• decision logs

• no ethics body
• weak charter
• missing minutes

Implement data accountability practices including data quality, lineage, and governance

• data management framework
• data quality records
• lineage documentation
• review cadence

• no data framework
• weak quality controls
• missing lineage

Ensure AI model design, selection, and training is governed by responsible practices

• algorithm design documentation
• training procedure records
• model versioning
• approval logs

• no design documentation
• weak versioning
• missing approvals

Per MAF: operations. Requirements include (a) Data Management + Data Quality and Provenance for Training + (b) Algorithm Design and Training + (c) safety + reliability + robustness + (d) Third-Party AI Component Due Diligence.

• SG AI Gov evidence for SGAIGOV-3

• AI ethics body + GenAI provenance partial

Implement measures to minimise inherent biases in data collection and pre-processing

• bias assessment reports
• data sampling documentation
• mitigation records
• monitoring dashboards

• no bias assessment
• weak mitigation
• missing dashboards

Ensure AI decision-making processes can be explained in understandable terms relative to the risk level

• explainability technique documentation
• SHAP or LIME outputs
• explanation interfaces
• user testing records

• no explanation techniques
• weak interfaces
• no user testing

Ensure AI systems produce consistent and traceable outcomes for review and audit

• reproducibility documentation
• model versioning
• training trace records
• audit trail

• weak reproducibility
• no versioning
• missing trace

Continuously monitor and tune AI models to detect and correct drift, bias, and performance degradation

• monitoring dashboards
• tuning records
• drift detection alerts
• retraining triggers

• no monitoring
• weak tuning records
• missing drift detection

Per MAF: risk + human involvement. Requirements include (a) risk assessment for AI applications + (b) Human-in-the-Loop or Human-over-the-Loop as appropriate + (c) determine level of human involvement per risk + (d) safety + reliability + robustness.

• SG AI Gov evidence for SGAIGOV-2

• AI ethics body + GenAI provenance partial

Per MAF: stakeholder interaction. Requirements include (a) transparency to users + (b) explainability + (c) communication of AI use + (d) recourse mechanisms + (e) user awareness.

• SG AI Gov evidence for SGAIGOV-4

• AI ethics body + GenAI provenance partial

Be transparent about the use of AI in products and services through disclosure to stakeholders

• public transparency documents
• AI use disclosures
• stakeholder communications
• policy documents

• no transparency documents
• weak disclosures
• missing communications

Communicate AI-related policies in plain and accessible language to relevant stakeholders

• plain-language communications
• accessibility-compliant materials
• multilingual records
• user research

• jargon-heavy materials
• no accessibility
• missing multilingual

Provide accessible channels for individuals to raise concerns or provide feedback on AI-driven decisions

• feedback channels documentation
• complaint procedures
• response records
• improvement tracking

• no feedback channels
• weak complaint procedures
• missing improvement tracking

Proactively notify individuals when AI is used to make decisions that significantly affect them

• AI use disclosure standards
• user notification templates
• consent records
• compliance dashboards

• no AI disclosure
• weak notification design
• missing consent