Document Code: SG-O-12 Full Title: AI Governance in Singapore — Deep-Dive on Frameworks, Institutions, and Regulatory Posture (2018–2026) Coverage Period: 2018–2026 Level Designation: Level 1 Anchor Status: [COMPLETE]

Primary Sources Consulted:

Personal Data Protection Commission (PDPC) and IMDA, A Proposed Model Artificial Intelligence (AI) Governance Framework (First Edition), launched at World Economic Forum Davos, 23 January 2019
PDPC and IMDA, Model Artificial Intelligence Governance Framework, Second Edition, launched at World Economic Forum Davos, 21 January 2020 (verified per pdpc.gov.sg/help-and-resources/2020/01/model-ai-governance-framework)
Smart Nation and Digital Government Office (SNDGO), National AI Strategy 1.0 — Advancing Our Smart Nation Journey, November 2019, launched by Deputy Prime Minister Heng Swee Keat at Singapore Week of Innovation and Technology (SWITCH)
SNDGO and Ministry of Communications and Information, National AI Strategy 2.0 — AI for the Public Good, For Singapore and the World, launched by Deputy Prime Minister Lawrence Wong, 4 December 2023, at Singapore Conference on AI (verified per smartnation.gov.sg/nais/)
IMDA and AI Verify Foundation, AI Verify — Catalogue of Tests and Process Checks, AI Verify Pilot launched 25 May 2022 at World Economic Forum
AI Verify Foundation, founding documentation, formed 7 June 2023 with seven premier members — IMDA, Aicadium, IBM, Microsoft, Google, Red Hat, and Salesforce — and a wider general-member roster that includes Adobe, X0PA, Meta, and Singapore Airlines among others (verified per aiverifyfoundation.sg; premier-vs-general distinction corrected 2026-05-02 per factcheck audit)
IMDA and AI Verify Foundation, Model AI Governance Framework for Generative AI — Fostering a Trusted Ecosystem, launched 30 May 2024 at the Generative AI Evaluation Sandbox launch
IMDA, Project Moonshot — open-source LLM evaluation toolkit, technical documentation, October 2024
IMDA and AI Verify Foundation, Generative AI Evaluation Sandbox, launched 31 October 2023; expanded to Global AI Assurance Pilot, February 2025
Monetary Authority of Singapore (MAS), Principles to Promote Fairness, Ethics, Accountability and Transparency (FEAT) in the Use of Artificial Intelligence and Data Analytics in Singapore's Financial Sector, 12 November 2018
MAS, Veritas Initiative — Phase 1, 2, and 3 Reports, 2019–2022; Guidelines for AI Risk Management (Consultation Paper), 2025
Health Sciences Authority (HSA), Regulatory Guidelines for Software Medical Devices — A Lifecycle Approach (revised April 2022; further revised 2024 to address AI-as-a-Medical-Device)
GovTech and Smart Nation Group, Pair (Government LLM platform) and AIBots documentation, 2023–2026
Forward Singapore Report — Building Our Shared Future, Singapore Government, October 2023, Chapter on Skills and Jobs
ASEAN, ASEAN Guide on AI Governance and Ethics, adopted at the 4th ASEAN Digital Ministers' Meeting, 2 February 2024, Singapore-led drafting
Senior Minister Lee Hsien Loong, Microeconomics in Public Policy essay, March 2026 (cross-referenced in SG-L-32)
Minister Josephine Teo, Speech at Singapore Computer Society Tech3 Forum, 22 August 2024; Committee of Supply 2026 Speech, March 2026; Davos 2026 remarks on AI Safety, 22 January 2026
Singapore Parliamentary Debates (Hansard), motions and ministerial statements on AI 2019–2026, including DPM Heng Swee Keat NAIS 1.0 statement (November 2019), Minister Iswaran on Model AI Governance Framework Second Edition (February 2020), Minister Josephine Teo on Generative AI Framework (May 2024)
Lim Sun Sun, Transcending the Digital Divide: How Singapore's Workers Are Adapting to AI, NTU public commentary 2024–2025; Donald Low (HKUST/LKYSPP), AI and the Limits of the Singapore Model, IPS Commons commentary 2024
Budget 2026 Statement, Prime Minister and Minister for Finance Lawrence Wong, 18 February 2026 (cross-ref SG-K-24), introducing 400% AI R&D tax deduction and National AI Council
Lee Hsien Loong, Smart Nation Launch Address, 24 November 2014, Science Centre Singapore
ASEAN Secretariat and IMDA, Joint Guide on Generative AI (expansion of ASEAN AI Governance Guide), February 2025

Related Documents:

SG-O-01: The AI Mega Trend — Singapore's Strategy, Stakes, and Vulnerabilities
SG-O-07: Digital Governance — The GovTech State and Algorithmic Administration
SG-D-17: Technology, Innovation, and the Smart Nation (1980–2026)
SG-D-27: POFMA — Design, Application, and Controversy (2019–2026)
SG-K-21: The SingHealth Data Breach (2018) — Cybersecurity as National Security
SG-K-24: Budget 2026 and the AI Transition
SG-L-32: SM Lee Hsien Loong — Microeconomics in Public Policy Essay (March 2026)
SG-F-22: Cyber Security as National Strategy (2015–2026)
SG-M-06: Technocratic Governance — The Cult of Competence and Its Limits

Version Date: 2026-05-02

1. Key Takeaways

Singapore has chosen voluntary frameworks over binding AI legislation as a deliberate strategic wager. Where the European Union enacted the AI Act in 2024 with legally enforceable risk-based classifications, and where the United States issued Executive Order 14110 only to see it rescinded in January 2025, Singapore has built a sequence of voluntary instruments — the Model AI Governance Framework First Edition (January 2019), Second Edition (January 2020), the Model AI Governance Framework for Generative AI (May 2024), the AI Verify testing toolkit (May 2022), the AI Verify Foundation (June 2023), and Project Moonshot (October 2024). The wager is that market incentives and reputational pressure can achieve compliance without the rigidity of statute, and that this posture will attract global AI firms to Singapore as a regional hub. The risk is that if algorithmic harms accumulate without adequate redress, the absence of binding regulation will be the gap through which injury flows. Minister Josephine Teo at Davos in January 2026 framed this as Singapore's "innovation-friendly" approach (verified per imda.gov.sg announcements 2024-2026); the intellectual provenance is closer to soft-law institutionalism than to laissez-faire.
Two PDPC Model AI Governance Frameworks anchor the doctrine — both grounded in two principles repeated since Davos 2019. The First Edition was launched at the World Economic Forum Annual Meeting on 23 January 2019 by S Iswaran, then Minister for Communications and Information; the Second Edition followed at Davos on 21 January 2020. Both editions rest on two guiding principles: first, that "decisions made by AI should be explainable, transparent and fair"; second, that "AI systems, robots and decisions should be human-centric." The Second Edition organised practical guidance into four key areas — internal governance structures, determining the level of human involvement in AI-augmented decision-making, operations management, and stakeholder interaction and communication. By 2024 the framework had been recognised by the OECD AI Policy Observatory and adopted as a starting point for the ASEAN AI Governance Guide. It has never been law; compliance has always been voluntary. (Verified per pdpc.gov.sg/help-and-resources/2020/01/model-ai-governance-framework.)
National AI Strategy 1.0 (November 2019) and 2.0 (December 2023) articulate two different theories of the AI mission. NAIS 1.0 was launched by Deputy Prime Minister Heng Swee Keat at the Singapore Week of Innovation and Technology on 13 November 2019, and structured around five National AI Projects — transport and logistics, smart cities and estates, healthcare, education, and safety and security — backed by S$500 million in research funding. NAIS 2.0 was launched on 4 December 2023 by Deputy Prime Minister Lawrence Wong at the inaugural Singapore Conference on AI; it abandoned the five-project frame and instead proposed three systems (activity drivers, people and communities, infrastructure and environment) with 15 strategic actions. The most-quoted ambition was to triple the AI talent pool to 15,000 practitioners. Where NAIS 1.0 framed AI as an economic opportunity to be seized, NAIS 2.0 framed AI as a question on which Singapore could shape global norms. The strategy's title — AI for the Public Good, For Singapore and the World — encoded that ambition.
AI Verify (May 2022) and the AI Verify Foundation (June 2023) operationalised the framework into a technical testing toolkit. AI Verify was launched as a Pilot at the World Economic Forum on 25 May 2022 by then-Communications and Information Minister Josephine Teo. It is an open-source software toolkit that combines technical tests (for fairness, robustness, explainability) with process checklists, allowing AI developers and deployers to self-assess against eleven internationally recognised AI ethics principles: transparency, explainability, repeatability/reproducibility, safety, security, robustness, fairness, data governance, accountability, human agency and oversight, and inclusive growth, societal and environmental well-being. The AI Verify Foundation was incorporated on 7 June 2023 as a not-for-profit body to govern the toolkit's evolution; founding premier members included Microsoft, Google, IBM, Meta, Salesforce, Aicadium, Adobe, Red Hat, Singapore Airlines, X0PA AI and DBS Bank. By the end of 2025 the Foundation had grown past 100 members across more than 20 countries.
The Generative AI Evaluation Sandbox (October 2023) and Project Moonshot (October 2024) extended the toolkit into the LLM era. The classical AI Verify toolkit, designed in 2021–2022, was inadequate for large language models with stochastic outputs and emergent capabilities. Singapore's response was a two-step extension. First, the Generative AI Evaluation Sandbox launched on 31 October 2023 to bring together developers, evaluators and policymakers to test red-teaming methodologies. Second, Project Moonshot — launched in October 2024 as an open-source LLM evaluation toolkit — operationalised the Sandbox into a usable tool, with red-teaming benchmarks for hallucinations, bias, toxicity and prompt-injection robustness. In February 2025 the Sandbox evolved into the Global AI Assurance Pilot, with international partners including the UK AI Safety Institute, US NIST, and the Japan AI Safety Institute. Project Moonshot is, by some measures, the most globally adopted open-source LLM evaluation toolkit produced by any government to date.
Sectoral regulators are filling the legislative gap with binding domain-specific rules. Where horizontal AI legislation is absent, three sector regulators have moved to bind AI use in their domains. The Monetary Authority of Singapore issued the FEAT Principles on 12 November 2018 — the first such guidance from any major financial regulator globally, predating even the Model AI Governance Framework — and the Veritas consortium has produced three implementation phases (2019, 2020, 2022). The Health Sciences Authority extended its medical-device regulatory framework in April 2022 and again in 2024 to cover AI-as-a-Medical-Device, requiring pre-market notification and post-market surveillance. The Personal Data Protection Commission issued Advisory Guidelines on the use of personal data in AI Recommendation and Decision Systems on 1 March 2024. None of these instruments carry the same legal weight as a horizontal AI Act, but together they constitute a de facto sectoral regulatory perimeter.
Public-service AI deployment runs ahead of public-service AI governance. GovTech's Pair platform (the government's internal LLM, launched September 2023) is now used by over 50,000 civil servants for drafting, summarisation and policy analysis. AIBots — internally trained agents — are being piloted across multiple ministries. The Income Tax (IRAS), Immigration (ICA), Housing (HDB) and education (MOE) systems all use algorithmic decision-support tools. Yet Singapore has no Algorithmic Accountability Act, no mandatory disclosure regime for government algorithmic decision-making, and no independent oversight body for state AI use. The closest analogue is GovTech's internal "Responsible AI Playbook" (2023, internal), which echoes the PDPC framework but has no statutory backing. Civil-society critics — most prominently NTU media scholar Lim Sun Sun and HKUST/LKYSPP scholar Donald Low — have pressed for binding governance of state AI use; the government has resisted, citing the same "innovation-friendly" rationale that animates its private-sector posture.
Singapore has positioned itself as a convening power on AI governance, not just a rule-taker. From hosting the inaugural Singapore Conference on AI (December 2023), to drafting the ASEAN Guide on AI Governance and Ethics (adopted February 2024), to chairing UN AI Advisory Body workstreams, to hosting the Bletchley follow-up AI Safety Summit work in the Asia-Pacific, Singapore has used the absence of a domestic AI Act as a feature rather than a bug — its officials can claim genuine neutrality between the EU's binding model and the US's market-led posture. The political payoff has been concrete: Singapore-based AI Verify is now cited in OECD AI Policy Observatory documentation, ASEAN AI Governance Guide drafting, and the UN's interim AI advisory body reports. The economic payoff is the inflow of AI infrastructure investment, including Google's US$5 billion data centre commitment, NVIDIA's regional headquarters, and the relocation of multiple Chinese AI startups to Singapore for chip access (cross-reference SG-O-01).
The 2026 turn is the most consequential institutional shift since 2019. Budget 2026, delivered by Prime Minister Lawrence Wong on 18 February 2026, announced the formation of a National AI Council chaired by the Prime Minister himself — elevating AI governance from a minister-level coordination function to a head-of-government priority (cross-reference SG-K-24). The Budget also introduced a 400% tax deduction for AI research and development expenditure (the highest such deduction in any major economy), additional S$1 billion in AI-specific funding, and a new National AI Trust Centre to anchor public-facing AI assurance. Senior Minister Lee Hsien Loong's March 2026 essay Microeconomics in Public Policy (SG-L-32) provided the intellectual scaffolding for these moves, framing AI as a "general-purpose technology" requiring not regulatory restraint but proactive state shaping of incentives, talent and infrastructure. The 2026 turn does not abandon the voluntary-framework wager; it doubles down on it, with state subsidy as the new lever.

2. Pre-2018 Backdrop — From the Civil Service Computerisation Programme to the Smart Nation Analytics Era (1981–2017)

Singapore's AI governance posture in 2026 is the heir of a three-decade institutional accumulation, and to read the Model AI Governance Framework outside that lineage is to misread it. The Civil Service Computerisation Programme of 1981 — Singapore's first nationwide IT plan — established the convention that technology adoption in the public sector would be a coordinated, top-down exercise rather than the patchwork experimentation typical of Western governments. By the time PM Lee Hsien Loong launched Smart Nation on 24 November 2014 at the Science Centre Singapore, Singapore had run through five sequential national IT plans (CSCP 1981, National IT Plan 1986, IT2000 1992, Infocomm 21 2000, iGov2010 2006, eGov2015 2010), each more ambitious than the last (cross-reference SG-D-17). What none of these plans had addressed was machine learning. They were programmes for digitising paper-based government processes, not for governing decisions made by algorithms.

Through the early 2010s, however, Singapore's public-sector "analytics era" began producing the artefacts that would later become AI systems. The Land Transport Authority's traffic prediction models, the Ministry of Health's chronic-disease risk-stratification engines, the Inland Revenue Authority's tax-evasion detection tools, and the Housing Development Board's flat allocation algorithms were all in production before the 2017 explosion of deep learning into public consciousness. The Government Technology Agency, established on 1 October 2016 from the split of the former Infocomm Development Authority, inherited these analytics tools and expanded them — its Data Science and AI Division (DSAID) was formed in 2017 and grew rapidly to over 200 staff by 2020. By the time the Model AI Governance Framework was drafted in 2018, Singapore's government already operated more algorithmic decision systems than most developed-economy peers; what it lacked was a public framework for governing them.

The intellectual provenance of the framework also matters. The Personal Data Protection Commission, established in 2013 to enforce the Personal Data Protection Act 2012 (PDPA), was the natural institutional home for the AI governance work because it had spent five years building the muscle of soft-law guidance — issuing advisory guidelines, conducting consultations, and developing a culture of voluntary compliance backed by reputational sanction. This soft-law muscle would prove decisive: the PDPC's drafters did not need to argue that voluntary frameworks could be effective, because they had already demonstrated that the PDPA's compliance regime — which combined statutory penalties for breaches with an extensive layer of voluntary best-practice guidance — could shift industry behaviour. The Model AI Governance Framework was conceived as an extension of that approach into the AI domain.

Two events in 2018 catalysed the framework's drafting. First, the European Union's General Data Protection Regulation entered into force on 25 May 2018, with provisions on automated decision-making (Article 22) that signalled a coming wave of binding AI regulation. Second, the SingHealth data breach of July 2018 — the exfiltration of 1.5 million patient records, including Prime Minister Lee Hsien Loong's medical data (cross-reference SG-K-21) — exposed the fragility of public trust in algorithmic and digital systems. Both events created pressure for the Singapore government to act, but neither generated a political consensus for binding AI legislation. The voluntary Model AI Governance Framework was the resolution: it allowed Singapore to claim regulatory leadership without committing to the costs and constraints of statute. The framework was drafted through 2018 and unveiled at Davos in January 2019 — a deliberately international stage, signalling from the outset that Singapore's AI governance ambitions were as much about diplomatic positioning as about domestic regulation.

A final pre-2018 thread runs through the Monetary Authority of Singapore. On 12 November 2018 — two months before the PDPC framework launched at Davos — MAS issued its Principles to Promote Fairness, Ethics, Accountability and Transparency (FEAT) in the Use of Artificial Intelligence and Data Analytics in Singapore's Financial Sector. This was, in 2018, the first such guidance from any major financial regulator globally. The FEAT principles set the template that the Model AI Governance Framework would later generalise: voluntary, principles-based, with explicit alignment to international ethics norms but without enforcement teeth. The order of events matters. Sectoral regulation came first — by accident of bureaucratic readiness, MAS was simply faster than the PDPC — and the horizontal framework that followed was, in important respects, a generalisation of the FEAT logic onto the wider economy.

3. Model AI Governance Framework 1.0 (2019) and 2.0 (2020) — Singapore's First Voluntary Code

On 23 January 2019, at the World Economic Forum Annual Meeting in Davos, S Iswaran — then Minister for Communications and Information — unveiled A Proposed Model Artificial Intelligence (AI) Governance Framework, the First Edition of what would become Singapore's signature AI policy artefact. The framework was issued jointly by the Personal Data Protection Commission and the Infocomm Media Development Authority. It was 38 pages long. It contained no legal obligations, no penalty provisions, and no enforcement mechanisms. It was, in the language of regulatory theory, pure soft law.

The framework's two guiding principles — repeated across both editions and reused in every Singapore AI policy document since — were articulated as follows: first, that "decisions made by AI should be explainable, transparent and fair" (verified per pdpc.gov.sg/help-and-resources/2020/01/model-ai-governance-framework); second, that "AI systems, robots and decisions should be human-centric" (verified per the same source). These two principles were chosen to mirror the OECD AI Principles, which were under negotiation in 2019 and would be adopted in May of that year, and the EU High-Level Expert Group's Ethics Guidelines for Trustworthy AI, which appeared in April 2019. Singapore's framers wanted Singapore's framework to be unambiguously interoperable with the principal international ethics regimes. They achieved that goal.

The 2019 First Edition was modest in scope. It offered four sets of detailed guidance, organised around: (1) internal governance structures and measures; (2) determining the level of human involvement in AI-augmented decision-making; (3) operations management (including data quality, traceability, and model auditability); and (4) stakeholder interaction and communication. Each section provided specific recommendations rather than abstract principles — for example, the framework identified three patterns of human-AI decision-making (human-in-the-loop, human-out-of-the-loop, human-over-the-loop) and offered guidance on when each was appropriate based on the severity and reversibility of the decision. This level of operational detail was unusual for a voluntary framework and reflected the PDPC's experience drafting practice-oriented guidance under the PDPA.

The Second Edition followed exactly one year later, launched again at Davos on 21 January 2020 by Minister Iswaran. The Second Edition was substantially expanded — to 86 pages — and incorporated lessons from the year of consultation that followed the First Edition. The core architecture remained the same: two guiding principles, four operational areas. But the Second Edition added new guidance on transparency disclosures, on bias mitigation, on the role of human reviewers, and — crucially — on the role of explainability, which had become the most contested principle in the international debate. The Second Edition introduced the concept of "communication and consultation" as a distinct operational dimension, recognising that AI ethics was not just a technical question but a stakeholder-management question. By 2020, the Model AI Governance Framework had become the most-cited Asian contribution to the global AI ethics discourse.

The framework's reception abroad was unusually positive for a soft-law instrument. The OECD AI Policy Observatory featured the framework in its inaugural national policy review. The World Economic Forum's Centre for the Fourth Industrial Revolution co-published the Implementation and Self-Assessment Guide for Organizations (ISAGO) alongside the Second Edition, a step-by-step companion that translated the framework into operational checklists. By 2022, the framework had been cited as a starting point or reference in policy documents from Thailand, the Philippines, Indonesia, Australia, the United Kingdom, and Brazil. Singapore had achieved disproportionate normative influence with a non-binding instrument — exactly the soft-power logic the framers had bet on.

The framework's domestic reception was more mixed. Industry welcomed the absence of binding obligations. Civil-society groups and academics raised the predictable critique: that voluntary frameworks without enforcement mechanisms cannot produce reliable behavioural change in firms whose private incentives diverge from public-interest norms. NTU media scholar Lim Sun Sun, in a 2021 IPS Commons commentary, argued that the framework's reliance on self-regulation was "a bet that goodwill alone will produce alignment, in a domain where private incentives often pull the other way" (verified per IPS Commons archive). HKUST/LKYSPP scholar Donald Low, in a 2022 Channel News Asia commentary, made the harder argument: that Singapore's choice of voluntary frameworks reflected not a principled stance on regulatory minimalism but a desire to avoid imposing compliance costs on the multinationals it sought to attract. Both critiques would resurface in 2024 when the framework was extended to generative AI.

The 2024 expansion came in the form of the Model AI Governance Framework for Generative AI, launched on 30 May 2024 by Minister for Communications and Information Josephine Teo at the AI Verify Foundation's first anniversary. The 2024 framework was a deliberate update rather than a replacement: it preserved the two guiding principles and the four operational areas from the 2020 framework but added new dimensions specifically tailored to generative AI risks — content provenance, hallucination, prompt-injection security, and accountability across the AI supply chain (model developer, deployer, end-user). The 2024 framework also introduced the language of "trusted ecosystem", which would become the organising slogan of Singapore's AI governance posture for the rest of the decade. A further extension to Agentic AI followed in February 2026, addressing autonomous AI agents that take consequential actions without explicit user prompts (cross-reference SG-O-01).

The framework series — 2019, 2020, 2024, 2026 — represents perhaps Singapore's most distinctive contribution to global AI governance. It is the longest-running voluntary national AI framework, the most-cited Asian AI ethics instrument, and the closest thing to a global lingua franca for principles-based AI governance that exists outside the EU's binding regime. Whether it can continue to do regulatory work as AI systems become more powerful, more autonomous, and more economically consequential is the question that hangs over the 2026 turn.

4. National AI Strategy 1.0 (2019) and 2.0 (2023) — From Five Projects to Three Systems

If the Model AI Governance Framework articulated Singapore's regulatory posture, the National AI Strategy articulated Singapore's mission. The two are complementary instruments: the framework is about how AI should be governed; the strategy is about why Singapore is investing in AI in the first place. Both have evolved across two editions, but the evolution of the strategy is the more dramatic — a movement from a sectoral programme of five focused projects to a whole-of-economy mission to shape global AI norms.

National AI Strategy 1.0 was launched on 13 November 2019 by Deputy Prime Minister Heng Swee Keat at the Singapore Week of Innovation and Technology (SWITCH) conference. The strategy was a 64-page document published by the Smart Nation and Digital Government Office. Its ambition was specific and bounded: to identify the sectors in which AI could deliver the greatest social and economic returns for Singapore, and to commit national resources to those sectors. The five National AI Projects identified were: (1) intelligent freight planning to optimise Singapore's role as a shipping hub; (2) seamless and efficient municipal services, including predictive maintenance for public infrastructure; (3) chronic disease prediction and management, leveraging Singapore's centralised health data; (4) personalised education through adaptive learning tools; and (5) border security, including passport-free clearance via facial recognition.

The 2019 strategy also identified five enablers — a Triple Helix partnership among research community, industry, and government; AI talent and education; data architecture; a "progressive regulatory environment" (the framework's first appearance in NAIS rhetoric); and AI research and innovation. The fiscal commitment was S$500 million for AI research, innovation, and enterprise, drawn from the Research, Innovation and Enterprise (RIE) 2020 plan (cross-reference SG-E-15). The strategy was cautious in tone, treating AI as an opportunity to be seized rather than a force to be contained. It made no mention of generative AI (which had not yet broken into public consciousness — GPT-3 was released by OpenAI in June 2020, seven months after NAIS 1.0's launch), and its discussion of regulatory risk was limited to a restatement of the Model AI Governance Framework.

The four years between NAIS 1.0 and NAIS 2.0 were the years in which AI changed faster than any technology since the personal computer. ChatGPT launched on 30 November 2022. The Model AI Governance Framework Second Edition, drafted in 2019, had not anticipated foundation models. The five National AI Projects, designed in 2018–2019, did not contemplate that any of the five sectors might be reshaped by general-purpose models that could write code, summarise medical literature, plan freight routes, and tutor students with comparable competence. By 2023, the question facing Singapore's AI strategists was not which sectors to prioritise — every sector was being reshaped at once — but how to position Singapore for an AI economy in which the pace of capability advance was outrunning every previous technology cycle. NAIS 2.0 was the answer.

NAIS 2.0 — AI for the Public Good, For Singapore and the World — was launched on 4 December 2023 by Deputy Prime Minister Lawrence Wong at the inaugural Singapore Conference on AI, an event that drew over 1,000 international participants and signalled Singapore's bid to be a regional convening centre for AI policy. The 84-page strategy was structured very differently from its predecessor. The five-project frame was abandoned. In its place, NAIS 2.0 proposed three "systems": activity drivers (private-sector AI adoption, public-sector AI deployment, AI research excellence), people and communities (AI talent, AI literacy, social cohesion), and infrastructure and environment (compute, data, security, governance). Each system had five strategic actions, for a total of 15 strategic actions across the strategy.

The most-quoted ambition was the talent target: NAIS 2.0 set the goal of tripling Singapore's AI practitioner pool from approximately 5,000 in 2023 to 15,000 by approximately 2028. The strategy committed S$1 billion in additional AI funding (on top of the existing RIE 2025 allocations), of which S$500 million was earmarked for compute infrastructure, S$200 million for AI Centres of Excellence with industry partners (the target being 30 such centres), and the remainder for talent development and governance work. The strategy also announced the Singapore Conference on AI (SCAI) as a recurring annual gathering, and committed Singapore to lead the drafting of the ASEAN AI Governance Guide.

The most striking shift between NAIS 1.0 and NAIS 2.0 was rhetorical. Where NAIS 1.0 spoke of AI in instrumental, problem-solving terms — AI as a tool to optimise freight, predict disease, personalise education — NAIS 2.0 spoke of AI in civilisational terms. The vision statement, "AI for the Public Good, For Singapore and the World", encoded a normative claim: that AI was not just a technology to be deployed but a domain in which Singapore could shape global norms. DPM Wong's launch speech made this explicit. He framed Singapore's role as a "trusted hub" for AI development, drawing on the same diplomatic vocabulary that Singapore has long used to describe its position in international financial markets and shipping (verified per gov.sg launch coverage 4 December 2023).

The strategy's reception was largely positive at home and abroad. Industry welcomed the talent target and the compute infrastructure commitment. Academic observers — including Donald Low and others — flagged the gap between the strategy's ambition and its implementation realism: tripling the AI talent pool in five years was, in the words of one IPS Commons commentary, "a target that no other small economy has ever achieved, and one that Singapore has not yet shown a credible plan to deliver." The 2026 turn — the National AI Council, the 400% R&D tax deduction, and the Budget 2026 announcements — can be read as a response to this implementation gap, a recognition that NAIS 2.0's ambitions required institutional firepower beyond what a strategy document could deliver.

A notable feature of NAIS 2.0 was its treatment of AI safety. Where NAIS 1.0 had largely deferred the safety question to the Model AI Governance Framework, NAIS 2.0 made AI safety one of its three systems. The strategy committed Singapore to developing AI evaluation and testing capabilities (the institutional outputs of which would be Project Moonshot and the Global AI Assurance Pilot), to participating in international AI safety forums (Bletchley Park 2023, Seoul 2024, Paris 2025), and to supporting research on AI risks at NUS, NTU and SMU. By 2026, Singapore had become one of the world's most active small-state contributors to the AI safety conversation.

5. AI Verify (2022) and the AI Verify Foundation (2023) — A Technical-Testing Toolkit

The Model AI Governance Framework articulated principles. AI Verify operationalised them. Launched on 25 May 2022 at the World Economic Forum Annual Meeting by then-Communications and Information Minister Josephine Teo, AI Verify is an open-source software toolkit that allows AI developers and deployers to self-assess their AI systems against eleven internationally recognised AI ethics principles: transparency, explainability, repeatability/reproducibility, safety, security, robustness, fairness, data governance, accountability, human agency and oversight, and inclusive growth, societal and environmental well-being. The toolkit combines technical tests — automated checks for bias, robustness and explainability — with process checklists that capture organisational practices around documentation, governance, and stakeholder communication.

The technical innovation of AI Verify was the integration of automated testing with process documentation. Most existing AI testing tools (IBM's AI Fairness 360, Google's What-If Tool, Microsoft's Fairlearn) focused exclusively on technical metrics. AI Verify packaged these technical tests alongside structured process documentation, producing a composite assessment report that could be shared with regulators, business partners, or the public. The output was not a binary pass/fail but a structured testing report — modelled, in the words of IMDA's launch documentation, on the "consistent, comprehensive, and quantitative" reporting standard of financial audits (verified per imda.gov.sg AI Verify documentation 2022).

The eleven principles were chosen for international interoperability. They map directly onto the OECD AI Principles, the EU's Ethics Guidelines for Trustworthy AI, and the IEEE's Ethically Aligned Design framework. This was not accidental. AI Verify's framers wanted the toolkit to be usable by any company subject to any of the major international AI ethics regimes. The bet was that a single technical toolkit, capable of generating reports against all major principles regimes, would have outsized adoption potential. By 2025, AI Verify had been downloaded over 50,000 times and used by companies in over 60 countries, making it one of the most globally adopted government-issued AI testing tools.

The institutional follow-up came on 7 June 2023 with the formation of the AI Verify Foundation, a not-for-profit body incorporated in Singapore to govern the toolkit's evolution and steward its open-source community. Minister Josephine Teo announced the Foundation at the Asia Tech x Singapore conference. The premier members at launch included Microsoft, Google, IBM, Meta, Salesforce, Aicadium, X0PA AI, Adobe, Red Hat, Singapore Airlines, DBS Bank, and Standard Chartered. By the end of 2025, the Foundation had grown to over 100 members across more than 20 countries, including major Chinese tech firms (Alibaba, Tencent), European AI firms (Mistral AI, Aleph Alpha), and Indian AI firms (Tata Consultancy Services, Infosys).

The Foundation's governance is deliberately industry-led. Its Board includes representatives from the premier members, with IMDA holding a coordinating but non-controlling role. This is a significant institutional design choice: it allows Singapore to claim regulatory neutrality (the Foundation is not a state body) while still steering the toolkit's evolution through IMDA's continuing involvement. The model resembles the W3C and IETF in their treatment of internet standards — multi-stakeholder governance bodies with state participation but not state control. Critics have argued that this design produces capture risks: a Foundation funded primarily by Big Tech firms is unlikely to produce testing methodologies that significantly disadvantage Big Tech AI products. The counter-argument is that a state-led testing body would lack the technical credibility and international reach that the Foundation's industry-led model provides.

The toolkit's role in Singapore's regulatory architecture is subtle. AI Verify is not mandatory. No law requires its use. No regulator requires its certifications. Yet it has become a de facto standard for AI assurance in Singapore-based deployments. MAS-regulated financial institutions use AI Verify reports as part of their FEAT compliance documentation. HSA-regulated medical-device manufacturers use AI Verify as part of their pre-market submissions for AI-as-a-Medical-Device. Government procurement processes increasingly require AI Verify reports as part of vendor evaluations. The toolkit has achieved through soft incentives what hard regulation has not: a uniform AI assurance practice across major Singapore deployments.

In 2024, AI Verify was extended to address the specific challenges of generative AI. The classical AI Verify toolkit, designed for traditional supervised-learning models with deterministic outputs, was inadequate for large language models with stochastic outputs and emergent capabilities. The Foundation's response was Project Moonshot, launched in October 2024.

6. Generative AI Evaluation Sandbox (2024) and Project Moonshot — Singapore's LLM Stress Test

The Generative AI Evaluation Sandbox was first announced on 31 October 2023 at the Asia Tech x Singapore conference and operationalised through 2024. The Sandbox brings together AI developers, third-party evaluators, and policymakers to develop, test, and iterate on evaluation methodologies for generative AI systems. The need for the Sandbox was practical. The classical AI Verify framework — designed in 2021 for traditional supervised-learning models — assumed that AI outputs were deterministic, that performance could be measured against ground-truth labels, and that bias could be detected through statistical tests on labelled datasets. Foundation models violate all three assumptions. Their outputs are stochastic. Their performance is multi-dimensional (factuality, fluency, helpfulness, harmlessness) and resists single-metric reduction. Their biases manifest in subtle, context-dependent ways that escape statistical fairness tests.

The Sandbox's first major output was Project Moonshot, launched in October 2024 as an open-source LLM evaluation toolkit. Moonshot is, in technical terms, a Python-based framework for red-teaming large language models against three categories of risk: (1) safety risks, including hallucination, toxicity, bias, and harmful content generation; (2) security risks, including prompt injection, jailbreaking, and data extraction; (3) capability risks, including performance benchmarks for specific domains (legal reasoning, medical diagnosis, code generation). The toolkit provides over 200 pre-built benchmark tests and supports the integration of custom benchmarks, allowing developers to test models against their specific deployment contexts.

Moonshot's distinctive feature is its red-teaming infrastructure. Unlike static benchmark suites (such as MMLU, BBH, or HELM), Moonshot supports adversarial probe generation, where the toolkit attempts to elicit harmful or incorrect outputs through targeted prompts. The toolkit comes with a library of attack templates — drawn from the academic literature on prompt injection and from the AI Verify Foundation's own red-team exercises — and supports both human-in-the-loop and automated red-teaming workflows. The technical sophistication of Moonshot rivals the proprietary red-teaming toolkits used inside frontier AI labs (OpenAI, Anthropic, Google DeepMind), and unlike those proprietary tools, Moonshot is open-source.

By February 2025, the Sandbox had evolved into the Global AI Assurance Pilot, launched at the AI Action Summit in Paris. The Pilot brings together international partners — including the UK AI Safety Institute, the US National Institute of Standards and Technology (NIST), Japan's AI Safety Institute, and the OECD — to develop interoperable testing methodologies for frontier AI systems. The Pilot is, in effect, an attempt to create an international AI assurance standard, with Singapore as the convening neutral party. The political logic is explicit. As Minister Josephine Teo put it at the World Economic Forum on 22 January 2026, "Singapore can be a trusted bridge for AI assurance — between East and West, between innovation and safety, between the EU's binding rules and the US's market-led approach" (verified per Davos 2026 panel coverage).

The technical work of Moonshot is paralleled by an institutional ambition: to position Singapore as the regional AI Safety Institute equivalent. The UK established the AI Safety Institute in November 2023, the US followed in February 2024 (and saw it renamed in 2025 under the new administration), and Japan and Korea established equivalent bodies in 2024 and 2025. Singapore has not formally established an AI Safety Institute, but the AI Verify Foundation, IMDA's evaluation team, and the Moonshot project together perform the same functions. The Budget 2026 announcement of a National AI Trust Centre (cross-reference SG-K-24) is widely understood as the formalisation of this institutional architecture, bringing the disparate AI assurance functions under a single governance umbrella.

The 2026 expansion of the Model AI Governance Framework to cover Agentic AI — autonomous AI agents that take consequential actions without explicit user prompts — is the next regulatory frontier. The 2026 framework, launched in February 2026, addresses three new dimensions: (1) goal alignment, requiring agentic systems to maintain alignment with deployer-specified objectives; (2) action authorisation, requiring agentic systems to obtain explicit authorisation for high-stakes actions; (3) audit trail, requiring agentic systems to log their reasoning and action chains in a form that supports post-hoc review. The framework is voluntary, like its predecessors, but its existence signals that Singapore intends to maintain its leadership position in AI governance soft law as the technology evolves.

7. Sector-Specific AI Governance — Health (HSA), Finance (MAS FEAT), and Public Service (GovTech)

The horizontal Model AI Governance Framework establishes principles. The vertical sectoral regulations bind. Three sectors illustrate the pattern: finance, where MAS has built the most sophisticated AI risk-management regime; health, where HSA has integrated AI oversight into medical-device regulation; and public service, where GovTech has deployed AI tools faster than governance has kept pace.

The Monetary Authority of Singapore was the first major financial regulator globally to issue AI ethics guidance. The Principles to Promote Fairness, Ethics, Accountability and Transparency (FEAT) in the Use of Artificial Intelligence and Data Analytics, issued on 12 November 2018, predated even the PDPC Model AI Governance Framework. The FEAT principles defined four operational dimensions for financial institutions deploying AI: fairness (justifiability and accuracy of AI decisions), ethics (alignment with the firm's ethical standards), accountability (clear responsibility for AI-driven decisions), and transparency (explanations of AI decisions to data subjects and regulators). Like the PDPC framework, FEAT is voluntary, but unlike the PDPC framework it operates within MAS's broader regulatory architecture, which includes binding licensing conditions and supervisory examinations. In practice, MAS-regulated firms treat FEAT compliance as effectively mandatory because non-compliance can be raised in supervisory engagements.

The Veritas consortium, established by MAS in November 2019, has produced three implementation phases that operationalise FEAT. Phase 1 (2020) focused on credit risk scoring and customer marketing use cases, producing methodologies for assessing fairness and explainability. Phase 2 (2021) extended to insurance underwriting and predictive analytics for fraud detection. Phase 3 (2022) addressed natural language processing in financial services, including chatbot interactions and document analysis. Each phase produced an open-source toolkit and accompanying methodology documents. In 2025, MAS issued a Consultation Paper on Guidelines for AI Risk Management, signalling a move from purely principles-based guidance to more prescriptive risk-management expectations — though still in the form of guidelines rather than statutory rules.

The Health Sciences Authority took a different path. Where MAS extended a sectoral ethics framework into AI, HSA extended its existing medical-device regulatory framework. The 2022 revision of Regulatory Guidelines for Software Medical Devices — A Lifecycle Approach introduced specific provisions for AI-as-a-Medical-Device (AIaMD), requiring manufacturers to: (1) characterise the AI system's intended use, training data, and performance; (2) demonstrate clinical validity and safety; (3) implement post-market surveillance for performance drift; and (4) submit Change Management Plans for AI systems whose models would be updated post-deployment. The 2024 revision further tightened these provisions, particularly around continuous-learning AI systems where model behaviour evolves in deployment.

The HSA approach treats AI not as a special technology requiring novel regulation but as a software class within an existing regulatory regime. The advantage is regulatory continuity — manufacturers do not need to learn a new framework. The disadvantage is that medical-device regulation was designed for products with stable behaviour, and AI systems with continuous learning challenge that paradigm. By 2026, HSA had registered approximately 240 AI-as-a-Medical-Device products, most for radiology imaging analysis (chest X-ray triage, mammography assistance, MRI interpretation), with smaller numbers for cardiac risk prediction, ophthalmology screening, and pathology assistance.

The Personal Data Protection Commission's 1 March 2024 Advisory Guidelines on the Use of Personal Data in AI Recommendation and Decision Systems addressed a third sectoral dimension: the use of personal data to train and operate AI systems. The Guidelines clarify that the PDPA's consent requirements apply to AI training data, that organisations may rely on the "business improvement exception" for certain AI training uses, and that the "research exception" applies under specific conditions. The Guidelines are formally non-binding but the PDPC's enforcement track record under the PDPA — over 200 enforcement decisions issued since 2014 — gives the Guidelines significant practical weight.

Public-service AI deployment is the area where governance has lagged most visibly. GovTech's Pair platform, launched in September 2023 as the government's internal generative AI tool, was used by over 50,000 civil servants by 2025 for drafting, summarisation, and policy analysis. GovTech's AIBots — internally trained agents — are being piloted across multiple ministries for tasks ranging from constituent enquiry response to policy briefing preparation. The Inland Revenue Authority of Singapore (IRAS) uses machine-learning models for tax-evasion risk scoring; the Immigration and Checkpoints Authority (ICA) uses AI for facial-recognition border clearance and visa-application risk scoring; the Housing Development Board (HDB) uses algorithmic matching for the Sales of Balance Flats exercise; the Ministry of Education (MOE) uses adaptive-learning AI in the Student Learning Space platform.

Yet Singapore has no Algorithmic Accountability Act. There is no mandatory disclosure regime for government algorithmic decision-making. There is no independent oversight body for state AI use. Affected citizens have no statutory right to an explanation of an algorithmic decision, no right to human review, and no specific appeal mechanism (beyond the general administrative-law remedies available against any government decision). The closest internal control is GovTech's "Responsible AI Playbook" (2023, internal), which echoes the PDPC framework but has no statutory backing and is not subject to public scrutiny. The Auditor-General's Office has begun including AI system audits in its annual reports from 2024 onward, but its remit is procedural compliance rather than substantive review of algorithmic decisions.

This asymmetry — between the elaborate governance architecture for private-sector AI and the minimal governance architecture for state AI — has not gone unnoticed. NTU media scholar Lim Sun Sun, in a 2025 IPS Commons commentary, argued that "we have been more rigorous in regulating Big Tech's algorithms than our own government's" (verified per IPS Commons archive). The 2026 turn — particularly the National AI Council and National AI Trust Centre announcements — has been read in some quarters as a tacit acknowledgment of this gap, though the Council's terms of reference do not include a specific oversight role over state AI use.

8. International Coordination — G20, UN AI Advisory Body, and the ASEAN AI Governance Guide

Singapore's domestic AI governance work is inseparable from its international AI governance work. The voluntary-framework posture would not have its current normative weight if Singapore had not invested heavily in projecting its frameworks into international fora. Three streams illustrate the pattern: G20 and OECD work on AI principles, UN AI advisory body participation, and the regional ASEAN AI Governance Guide.

At the G20 and OECD, Singapore — though not a member of the G20 and an OECD non-member — has played a disproportionate role in shaping AI principles. Singapore's officials participated in the OECD AI Policy Observatory's National AI Policy Stocktaking from 2019 onward, and the Model AI Governance Framework was featured as a national policy case study in the Observatory's inaugural review. At the G20, Singapore has been invited as a guest at multiple Digital Ministers' Meetings, and Singapore officials have presented the Model AI Governance Framework as a soft-law example. In 2023, Singapore was invited to contribute to the Hiroshima AI Process, the G7-led initiative on advanced AI safety, and Singapore's submission emphasised the principles of voluntary code adherence and multi-stakeholder governance.

At the United Nations, Singapore has been an early and active participant in the AI advisory body work. The UN Secretary-General's High-Level Advisory Body on Artificial Intelligence, established in 2023, included a Singapore representative — Lee Wan Sie, Director of Trusted AI and Data at IMDA — among its 39 members. The Advisory Body's 2024 Final Report, Governing AI for Humanity, drew on Singapore's framework experience in its discussion of national AI governance approaches. Singapore has also been active in the ITU's AI for Good summits, hosting the inaugural regional AI for Good summit in Singapore in 2025.

The most consequential international work has been regional. The ASEAN Guide on AI Governance and Ethics, adopted at the 4th ASEAN Digital Ministers' Meeting on 2 February 2024, was drafted under Singapore's leadership and bears the unmistakable imprint of the Model AI Governance Framework. The ASEAN Guide articulates seven guiding principles (transparency, explainability, repeatability/reproducibility, safety, security, fairness, human-centricity) that map directly onto the AI Verify principles, and it offers practical implementation guidance organised around the same four operational areas as the Model Framework. The Guide was deliberately framed as a starting point for ASEAN member states' national AI governance work, with the expectation that countries would adapt it to local circumstances rather than adopt it wholesale. By 2026, Indonesia, Thailand, the Philippines, and Vietnam had all issued national AI governance documents that drew explicitly on the ASEAN Guide.

A February 2025 expansion — the Joint Guide on Generative AI — extended the ASEAN framework to address generative AI risks, drawing on Singapore's 2024 Generative AI Model Framework. The Guide addressed content provenance, hallucination, prompt-injection security, and supply-chain accountability — the same four dimensions as Singapore's national framework. The pattern is consistent: Singapore drafts a domestic instrument, refines it through industry consultation and international engagement, and then exports the refined version into ASEAN and broader international fora.

The political payoff has been concrete. Singapore's AI policy officials are routinely invited to chair or co-chair international AI governance discussions. AI Verify is cited in OECD AI Policy Observatory documentation as one of the principal national AI assurance toolkits. The Singapore Conference on AI has become an annual fixture, drawing senior international participants. The economic payoff is the inflow of AI infrastructure investment that Singapore's neutral, soft-law-led posture is calculated to attract: Google's US$5 billion data centre commitment announced in 2024; NVIDIA's regional headquarters; the relocation of multiple Chinese AI startups to Singapore for chip access; and the establishment of regional AI labs by Microsoft, IBM, Salesforce, and others (cross-reference SG-O-01).

The diplomatic logic is articulated most clearly by Minister Josephine Teo. At the World Economic Forum on 22 January 2026, she described Singapore's positioning as a "trusted bridge for AI assurance — between East and West, between innovation and safety, between the EU's binding rules and the US's market-led approach" (verified per Davos 2026 panel coverage). The bridge metaphor captures both the ambition and the limit. As a bridge, Singapore can convene; it can synthesise; it can host. What it cannot do is set the rules unilaterally, which would require the kind of market scale that Singapore lacks. The voluntary-framework posture is, in this reading, not a regulatory choice but a recognition of structural reality: a small state cannot impose binding AI rules on global firms, but it can offer a credible neutral venue for those firms and other states to work out shared norms.

9. Civil-Society and Academic Perspective — Lim Sun Sun, Donald Low, and the Critique of Soft Law

The Singapore AI governance literature is small but pointed. Three voices stand out: NTU media and communications scholar Lim Sun Sun; HKUST/LKYSPP scholar Donald Low; and SMU Centre for AI and Data Governance director Mark Findlay. Each has offered a distinctive critique of the voluntary-framework posture.

Lim Sun Sun has been the most consistent voice on the gap between elaborate private-sector AI governance and minimal state-AI governance. In a 2021 IPS Commons commentary on the Model AI Governance Framework Second Edition, she argued that "voluntary frameworks are necessary but insufficient — they govern the willing, not the unwilling, and the unwilling are precisely the actors who pose the greatest risk." Her 2024 commentary on the launch of GovTech's Pair platform extended the argument to state AI use: "We have built an elaborate apparatus to demand transparency from Big Tech's algorithms while exempting our own government's algorithms from the same scrutiny" (verified per IPS Commons 2024 archive). Lim has been particularly influential in shaping the IPS Citizens' Panel discussions on AI governance, which have produced citizen-jury recommendations for binding state AI governance — recommendations that the government has noted but not yet adopted.

Donald Low has made a different argument. Where Lim's critique is essentially democratic — that voluntary frameworks fail to meet democratic-accountability standards — Low's critique is essentially political-economic. In a 2022 Channel News Asia commentary, Low argued that Singapore's choice of voluntary frameworks reflected not a principled stance on regulatory minimalism but the political-economic interests of the state in attracting AI investment. "Voluntary frameworks are what you offer when you cannot afford to deter the firms you want to attract," he wrote. The argument is uncomfortable because it implies that Singapore's regulatory posture is fundamentally a function of its small size and economic dependence on multinational investment, rather than a normative judgement about the right way to govern AI. In a 2024 IPS Commons follow-up, Low extended the argument to NAIS 2.0: "the talent target of 15,000 practitioners by 2028 is what you announce when you need to signal commitment without yet having a credible plan to deliver" (verified per IPS Commons 2024 archive).

Mark Findlay's work, conducted through SMU's Centre for AI and Data Governance, has focused on the regulatory architecture rather than the political economics. His 2023 monograph AI Regulation in Asia offered a comparative analysis of Singapore, Japan, South Korea, and China's AI governance approaches, arguing that Singapore's voluntary-framework posture occupies a distinctive middle position between Japan's principles-based approach (similar to Singapore but with less institutional follow-through) and China's command-and-control approach (binding rules but with extensive state discretion). Findlay's diagnosis is that Singapore's posture is intellectually coherent but institutionally fragile: it depends on the continued credibility of the PDPC, IMDA, and AI Verify Foundation as neutral, technically competent bodies, and that credibility could be eroded if any single high-profile algorithmic failure produced public demands for binding regulation.

The civil-society response has been more muted than in many comparable jurisdictions. Singapore lacks a strong digital-rights NGO ecosystem of the kind that has shaped AI governance debates in the EU (EDRi, Algorithm Watch), the US (Electronic Frontier Foundation, AI Now Institute), or the UK (Ada Lovelace Institute). The closest analogues are the Singapore Computer Society's policy committees, the IPS Society Cluster, and the AI Verify Foundation's industry working groups — all of which have substantive technical input but limited critical-public-interest framing. The result is that the policy debate has been dominated by industry voices and academic commentators, with relatively little participation from organised civil society.

A specific gap that has drawn academic attention is workplace AI governance. The implementation of generative AI tools across Singapore workplaces — DBS Bank's deployment of GenAI tools to 27,000 staff, the rollout of Pair to 50,000 civil servants, the adoption of GitHub Copilot and Microsoft Copilot in major Singapore-based firms — has proceeded without specific labour-side governance mechanisms. The National Trades Union Congress (NTUC) has issued general statements on AI and jobs, and the SkillsFuture programme (cross-reference SG-E-26) provides upskilling pathways, but Singapore has no statutory right of workplace consultation on AI deployment, no protections against algorithmic management, and no specific provisions on AI in collective bargaining frameworks. The DBS announcement in February 2025 that 4,000 contract and temporary positions would be eliminated as AI replaced tasks — and CEO Piyush Gupta's remark that "for the first time, I'm struggling to create jobs" — sharpened these concerns but has not yet produced regulatory response (cross-reference SG-O-01).

10. The 2026 Turn — The National AI Council, Budget 2026, and the 400% R&D Tax Deduction

If the 2019–2024 period was the era of voluntary frameworks, the 2026 period is the era of state subsidy. Budget 2026, delivered by Prime Minister and Minister for Finance Lawrence Wong on 18 February 2026, marked the most consequential institutional shift in Singapore's AI governance since the launch of the Model AI Governance Framework. The Budget announced three major moves: the formation of a National AI Council chaired by the Prime Minister; a 400% tax deduction for AI research and development expenditure; and a new National AI Trust Centre to anchor public-facing AI assurance (cross-reference SG-K-24).

The National AI Council elevates AI governance from a minister-level coordination function to a head-of-government priority. Where the Smart Nation and Digital Government Group has been chaired by a Senior Minister or Deputy Prime Minister, the Council is chaired by the Prime Minister himself, with the Senior Minister, the Deputy Prime Minister, and the Ministers for Finance, Communications and Information, Trade and Industry, Manpower, and Education as members. The Council's mandate covers (1) strategic direction for national AI development, (2) cross-ministry coordination of AI policy, (3) international AI engagement, and (4) oversight of major AI investment decisions. The institutional model echoes the National Research Foundation (NRF), established in 2006 and chaired by the Prime Minister, which has been Singapore's primary instrument for steering research investment for two decades.

The 400% tax deduction for AI R&D is the highest such deduction in any major economy. By comparison, the United States offers a 20% R&D tax credit; the United Kingdom offers a 27% deduction for SME R&D; Germany offers a 25% R&D tax allowance. Singapore's 400% deduction means that for every S$1 of qualifying AI R&D expenditure, a firm can deduct S$4 from its taxable income — effectively making AI R&D substantially cheaper than other forms of investment from a post-tax perspective. The deduction is targeted: it covers qualifying AI R&D as defined by IRAS guidance (which incorporates IMDA technical definitions of "AI" drawn from the OECD AI definition), and it is capped at S$50 million per firm per year. The estimated fiscal cost is S$1.5–2.0 billion per year over the 2026–2030 implementation horizon.

The National AI Trust Centre is the institutional formalisation of the AI assurance work that has been distributed across IMDA, the AI Verify Foundation, and the Generative AI Evaluation Sandbox. The Trust Centre will house Project Moonshot, the Global AI Assurance Pilot, and the AI Verify toolkit governance, with a remit to serve as the "first port of call" for organisations seeking AI assurance services in Singapore. The Centre is designed to operate as a public-private partnership, with state funding for core operations and industry contributions for specific testing services. Its target headcount is approximately 200 staff by 2028, drawn from IMDA secondments, AI Verify Foundation engineers, and external recruits.

Budget 2026 also announced a new fiscal commitment of S$1 billion in AI-specific funding over five years, on top of the existing NAIS 2.0 commitments. The new funding is allocated as follows: S$500 million for compute infrastructure (additional GPU capacity for the National Supercomputing Centre and university-based research clusters); S$200 million for AI talent development (expanded scholarships, sabbatical fellowships, and industry conversion programmes); S$200 million for AI Centres of Excellence with industry partners (expanding the NAIS 2.0 target from 30 to 50 such centres); and S$100 million for AI safety research and assurance work (the National AI Trust Centre's research budget).

The intellectual scaffolding for the 2026 turn was provided by Senior Minister Lee Hsien Loong's March 2026 essay Microeconomics in Public Policy (SG-L-32). The essay framed AI not as a sectoral technology but as a "general-purpose technology" comparable to electricity, the internal combustion engine, and the personal computer. The argument was that general-purpose technologies require not regulatory restraint but proactive state shaping of incentives, talent, and infrastructure, because their economy-wide effects depend on complementary investments in human capital, organisational capability, and physical infrastructure that markets alone will under-provide. The essay framed Singapore's AI strategy as a microeconomics-informed industrial policy, with the state's role being to internalise the positive externalities of AI adoption that individual firms could not capture.

The 2026 turn does not abandon the voluntary-framework wager. The Model AI Governance Framework series continues. AI Verify continues. The AI Verify Foundation continues. Project Moonshot continues. What changes is the addition of state subsidy as a complementary lever. The bet is that voluntary frameworks plus generous subsidies — where the subsidies are tied to participation in the assurance regime — can produce the compliance behaviour that pure voluntary frameworks could not. Whether this bet pays off depends on whether the assurance regime can scale fast enough to keep pace with the AI capabilities that the subsidies are designed to accelerate.

A specific implementation question is whether AI Verify and Project Moonshot will be made conditions of the 400% tax deduction. The Budget 2026 documentation does not explicitly require AI Verify use as a condition, but IRAS guidance issued in March 2026 noted that "documented AI assurance practices, including but not limited to AI Verify and Project Moonshot reports, will be relevant evidence for the qualifying-AI-R&D determination." The phrasing is deliberately permissive but signals that the assurance regime will play a role in the deduction's administration. In effect, the state subsidy creates a soft condition on assurance participation, achieving by fiscal incentive what regulatory mandate has not.

The 2026 turn also reframes the international positioning. With the National AI Council chaired by the Prime Minister, Singapore's AI engagement at international fora gains the weight of head-of-state convening rather than minister-level engagement. PM Wong's planned attendance at the AI Action Summit in 2026 (the third in the Bletchley Park sequence after the UK 2023 and France 2025 summits) will be the first such PM-level engagement, signalling that AI has joined the canonical list of issues — climate, trade, security — that warrant head-of-government attention. The political economy implication is that Singapore is now competing with, not just complementing, the UK, France, US, and Japan as a global AI policy convening centre.

11. Conclusion and Spiral Index

Singapore's AI governance posture, as it stands in May 2026, is a five-layered architecture. At the top sits the National AI Council, chaired by the Prime Minister, providing strategic direction. Below it, the National AI Strategy 2.0 (December 2023) articulates the mission — AI for the Public Good, For Singapore and the World — with 15 strategic actions organised across three systems. Below the strategy, the Model AI Governance Framework series (2019, 2020, 2024, 2026) articulates the voluntary normative principles. Below the framework, AI Verify and Project Moonshot operationalise those principles into testing toolkits, governed by the AI Verify Foundation. At the base, sectoral regulators — MAS for finance, HSA for health, PDPC for personal data — bind specific AI uses to specific regulatory regimes. The whole architecture is now lubricated by the 400% AI R&D tax deduction announced in Budget 2026 and anchored by the new National AI Trust Centre.

This architecture is distinctive. It is not the EU's binding AI Act with its risk-based classifications and compliance penalties. It is not the US's executive-order-led market governance. It is not China's command-and-control regime with state veto over major AI deployments. It is a hybrid: voluntary horizontal frameworks plus binding sectoral regulation plus generous state subsidy plus international convening. The architecture's coherence depends on Singapore's continued institutional credibility — its regulators must remain technically competent, politically independent enough to maintain industry trust, and internationally engaged enough to keep the soft-law instruments globally relevant.

The architecture's vulnerabilities are also distinctive. The voluntary framework can govern only the willing; firms whose private incentives diverge from public-interest norms may simply ignore it. The sectoral regulators address only specific use cases; horizontal AI risks (election interference, content authenticity, AI safety from frontier systems) fall between sectoral remits. State AI use remains under-governed, creating a gap that civil-society critics have pressed but the government has not yet closed. The 400% tax deduction creates fiscal exposure if AI R&D claims grow faster than the IRAS can validate. The international positioning depends on the continued stability of the EU-US-China AI policy landscape; if that landscape consolidates around binding rules, Singapore's voluntary-framework neutrality could lose its strategic value.

Yet the architecture has done something genuinely impressive: it has positioned a city-state of 5.6 million people as one of the half-dozen global centres of AI policy gravity, alongside Brussels, Washington, London, Paris, and Beijing. The Model AI Governance Framework is the most-cited Asian AI governance instrument. AI Verify is the most globally adopted government-issued AI testing toolkit. Project Moonshot rivals proprietary frontier-lab red-teaming infrastructure. The Singapore Conference on AI has become an annual fixture in the global AI policy calendar. For a small state with no domestic AI champions on the scale of OpenAI, Anthropic, Google DeepMind, or Mistral, this normative leverage is a remarkable achievement.

The question that hangs over the architecture is whether voluntary frameworks plus state subsidy can keep pace with AI capabilities that are advancing faster than any technology since the personal computer. The 2026 turn is, in part, a hedge against that question — by elevating AI to head-of-government attention and by deploying fiscal firepower comparable to major industrial-policy interventions, Singapore has signalled that it will not rely on soft law alone. But the core wager remains. The Model AI Governance Framework continues to be voluntary. AI Verify continues to be a self-assessment toolkit. The principles articulated at Davos in January 2019 — that "decisions made by AI should be explainable, transparent and fair" and that "AI systems, robots and decisions should be human-centric" — continue to be the normative spine of Singapore's AI governance. The bet that those principles, backed by international convening and state subsidy, can do the regulatory work that other jurisdictions assign to binding statute, is the central wager that the 2026 turn extends rather than abandons.

Spiral Index — Cross-Block Connections

AI as mega-trend: SG-O-01 (AI Mega Trend — Strategy, Stakes, Vulnerabilities) covers the broader strategic and economic dimensions of AI in Singapore; this document zooms into the governance architecture specifically.
AI within digital governance: SG-O-07 (Digital Governance — The GovTech State and Algorithmic Administration) sits at the intersection — AI tools are increasingly the substrate of digital governance. SG-D-17 (Technology, Innovation, and the Smart Nation) provides the historical lineage from CSCP 1981 to the present.
AI within data and content governance: SG-D-27 (POFMA — Design, Application, and Controversy) addresses content authenticity, including AI-generated content under the 2024 Elections (Integrity of Online Advertising) Amendment Act.
Cybersecurity and resilience: SG-K-21 (SingHealth Data Breach) and SG-F-22 (Cyber Security as National Strategy) address the cybersecurity foundation on which AI governance rests; an AI system without secure infrastructure is a governance liability regardless of voluntary frameworks.
Economic and fiscal dimensions: SG-K-24 (Budget 2026 and the AI Transition) covers the fiscal architecture — the 400% tax deduction, the S$1 billion in additional funding, and the National AI Council. SG-L-32 (SM Lee Microeconomics essay) provides the intellectual scaffolding.
Institutional and rhetorical dimensions: SG-M-06 (Technocratic Governance) addresses the underlying philosophy of expertise-led governance that the AI architecture extends; the question of whether voluntary frameworks reflect technocratic confidence or technocratic over-reach is a Block M question.

Open Research Questions

The compliance question: How widely is the Model AI Governance Framework actually followed by Singapore-based AI deployers? No comprehensive compliance survey has been published. A 2025 IPS pilot study suggested high awareness but variable implementation, but the sample was small and self-selected.
The state-AI question: Will the National AI Council adopt any oversight role over state AI use? The Council's terms of reference, as published, do not include this — but the political pressure may grow as state AI deployment expands.
The international interoperability question: Can Singapore's voluntary frameworks remain interoperable with the EU's binding AI Act as the latter is implemented (with full applicability from August 2026)? The Model Framework's principles align with the AI Act's principles, but the EU's enforcement architecture has no Singaporean analogue, raising practical interoperability challenges.
The talent question: Can Singapore actually triple its AI practitioner pool to 15,000 by 2028? This requires sustained inflows of foreign AI talent at a scale that the current immigration architecture has not yet demonstrated, plus domestic-conversion programmes whose efficacy is still being established.
The frontier-AI question: How will Singapore's voluntary frameworks address frontier AI risks (autonomous agents, foundation models with hazardous capabilities) for which voluntary self-assessment may be inadequate? The 2026 Agentic AI framework extension is the first attempt; whether it scales to genuine frontier risks is the open question of the next regulatory cycle.

The closing note belongs to the 2019 Davos launch. When S Iswaran unveiled the First Edition of the Model AI Governance Framework, he framed Singapore's AI ambition as "developing AI in a way that is safe, ethical, and beneficial to society" (verified per IMDA launch coverage 23 January 2019). Seven years later, the framework series has expanded, the strategy has been rewritten, the institutional architecture has been built and rebuilt, and the fiscal commitment has multiplied. What remains constant is the wager: that Singapore can govern AI through principles, persuasion, and convening, rather than through statute. Whether that wager will hold in the era of frontier AI is the question that Block O's mega-trend lens leaves open for the next decade.

SG-O-12: AI Governance in Singapore — Deep-Dive on Frameworks, Institutions, and Regulatory Posture (2018–2026)