The World’s First AI Governance Testing Toolkit
AI Verify represents Singapore’s most distinctive contribution to the global AI governance landscape. Launched by the Infocomm Media Development Authority (IMDA) on May 25, 2022, at the World Economic Forum in Davos, AI Verify is the world’s first AI governance testing framework and software toolkit that enables organizations to objectively assess and demonstrate the transparency, fairness, robustness, and accountability of their AI systems through standardized technical tests and process audits. The framework has been adopted by over 80 organizations across 15 countries, with 340 AI systems assessed through the toolkit by Q1 2026—establishing Singapore as the de facto standard-setter for practical AI governance in the Asia-Pacific region.
AI Verify emerged from a recognition that the AI governance discourse had become polarized between two inadequate positions: regulatory prescriptivism (exemplified by the EU’s AI Act, which imposes compliance obligations based on risk classification) and regulatory minimalism (exemplified by the U.S. approach under the previous administration, which relied primarily on voluntary industry commitments). Singapore’s assessment, articulated by IMDA Chief Executive Lew Chuen Hong in a foundational policy speech in March 2022, was that neither approach adequately served the needs of organizations deploying AI in production environments—prescriptive regulation imposed compliance costs without necessarily improving outcomes, while voluntary commitments lacked the verification mechanisms needed to build public trust.
AI Verify’s design philosophy occupies a pragmatic middle ground: it provides tools that organizations can use to demonstrate responsible AI practices, without mandating specific technical approaches or imposing blanket restrictions on AI use cases. The framework translates high-level AI ethics principles (drawn from the OECD AI Principles and Singapore’s own Model AI Governance Framework) into concrete, measurable tests that can be applied to real-world AI systems. This translation from principles to practice—what IMDA calls “operationalizing AI governance”—is the framework’s core innovation and the primary source of its international appeal.
Technical Architecture and Testing Methodology
AI Verify comprises three integrated components: the Governance Testing Framework (a structured assessment methodology), the Technical Toolkit (an open-source software package for automated testing), and the Report Generator (which produces standardized assessment reports for stakeholders).
The Governance Testing Framework evaluates AI systems across 11 governance dimensions derived from internationally recognized AI principles. Transparency measures the extent to which the AI system’s decision logic can be explained to affected stakeholders—assessed through technical tests of model interpretability (SHAP values, LIME explanations, attention visualization) and process audits of documentation practices. Fairness evaluates whether the AI system produces equitable outcomes across demographic groups—assessed through statistical tests of bias (disparate impact ratio, equalized odds, calibration across groups) applied to system outputs on representative test datasets. Robustness measures the AI system’s reliability under adversarial conditions and distributional shift—assessed through stress tests including adversarial perturbation, input corruption, and out-of-distribution detection.
Additional dimensions include accountability (governance structures for AI decision-making), safety (risk assessment and mitigation measures), human oversight (mechanisms for human review and override), data governance (training data quality, provenance, and consent), model lifecycle management (versioning, monitoring, and decommissioning procedures), privacy (data minimization, anonymization, and consent management), security (protection against model theft, data poisoning, and inference attacks), and inclusive design (accessibility and representativeness considerations).
The Technical Toolkit, published on GitHub under the Apache 2.0 license, provides automated testing capabilities for six of the eleven dimensions where quantitative measurement is feasible: transparency, fairness, robustness, data governance, privacy, and security. The toolkit supports AI systems built with major ML frameworks (PyTorch, TensorFlow, scikit-learn, XGBoost) and provides pre-built test suites for common AI application types including classification, regression, natural language processing, and computer vision. The toolkit processes approximately 35,000 test executions monthly through both the hosted service (operated by IMDA) and self-hosted installations at participating organizations.
The remaining five dimensions (accountability, safety, human oversight, model lifecycle management, and inclusive design) are assessed through structured process audits—questionnaires and evidence collection frameworks that evaluate organizational practices rather than technical model properties. The process audit methodology was developed in collaboration with PricewaterhouseCoopers (PwC), Deloitte, and Ernst & Young (EY), ensuring alignment with established audit practices familiar to enterprise compliance teams.
The Report Generator produces standardized assessment reports that communicate AI governance posture to multiple audiences. Technical reports provide detailed test results for AI engineering teams. Executive summaries present governance assessments in business language for leadership and board-level stakeholders. Public transparency reports provide citizen-accessible explanations of how an AI system was assessed and what the results indicate about its trustworthiness. The multi-audience reporting approach reflects Singapore’s recognition that AI governance must communicate effectively across organizational levels, not just to technical specialists.
AI Verify Foundation: Institutionalizing Governance
The AI Verify Foundation, established in June 2023 as an independent non-profit hosted by IMDA, provides the institutional home for the AI Verify framework’s ongoing development, community engagement, and international standardization efforts. The Foundation’s 22-member governing council includes representatives from technology companies (Google, IBM, Microsoft, Salesforce), financial institutions (DBS Bank, HSBC, Standard Chartered), academic institutions (NUS, MIT, Oxford), civil society organizations (Access Now, Partnership on AI), and government agencies (IMDA, the UK’s Alan Turing Institute, NIST).
The Foundation’s open-governance model enables international contribution to the framework’s evolution. Working groups comprising over 200 experts from 45 countries develop new testing methodologies, sector-specific assessment guides, and interoperability standards. The Generative AI Working Group, the largest and most active, has developed testing methodologies for large language models covering hallucination detection, bias in generated text, content safety filtering, and jailbreak resistance. These methodologies, published in the AI Verify GenAI Assessment Guide (March 2025), have been adopted by six AI companies for pre-deployment safety testing of their models.
The Foundation’s membership has grown to 120 organizations, with annual membership fees (ranging from SGD 5,000 for startups to SGD 100,000 for large enterprises) providing 60% of operating revenue. The remaining 40% comes from IMDA’s institutional support, which is structured to decrease annually as the Foundation achieves financial self-sustainability—a target date of 2028. The membership model creates an interesting governance dynamic: member organizations contribute to developing the standards by which their own AI systems will be assessed, requiring careful management of conflicts of interest through the Foundation’s Impartiality Policy.
Adoption Patterns and Sector-Specific Implementation
AI Verify adoption has followed distinct patterns across sectors, reflecting differences in regulatory pressure, risk exposure, and organizational AI maturity. Financial services leads adoption (38% of assessed systems), driven by MAS’s issuance of FEAT (Fairness, Ethics, Accountability, and Transparency) principles and the regulator’s explicit endorsement of AI Verify as an acceptable tool for demonstrating FEAT compliance. DBS Group has assessed 14 AI systems through AI Verify, including credit scoring models, anti-money laundering detection systems, and customer service chatbots, publishing summary results in its annual sustainability report.
Healthcare represents the second-highest adoption sector (22% of assessed systems), where AI Verify assessments serve both regulatory compliance and clinical trust-building purposes. The Ministry of Health’s AI in Healthcare Guidelines, published in 2024, recommend (but do not mandate) AI Verify assessment for clinical AI systems, with the assessment results required for submission in the Health Sciences Authority’s (HSA) medical device registration process for AI-based diagnostic tools. The National University Health System has assessed eight clinical AI systems, with fairness testing revealing and enabling correction of a demographic bias in a skin cancer detection model that underperformed for darker skin tones.
Government adoption, while not reflected in the public assessment statistics, has been significant. GovTech’s AI Safety Framework mandates internal governance assessments (using AI Verify methodologies adapted for government context) for all AI systems deployed in government services. The government’s Pair AI Platform, which provides generative AI capabilities to 45,000 civil servants, underwent an AI Verify assessment that evaluated content safety, bias in generated responses, and data privacy protections.
Enterprise adoption outside Singapore, while growing, has been concentrated in multinational corporations with Singapore operations that have extended AI Verify practices to their global AI portfolios. Notable international adopters include HSBC (assessing AI systems across its global operations), Siemens (applying AI Verify to industrial AI applications in manufacturing), and Unilever (using AI Verify for consumer-facing AI in marketing and supply chain optimization). The international adoption trajectory suggests that AI Verify could evolve from a Singapore-centric tool to a global standard, particularly as regulatory convergence creates demand for interoperable governance frameworks.
International Influence and Standards Alignment
AI Verify’s international influence extends beyond direct adoption to shaping the global AI governance standards landscape. The framework’s technical methodologies have been referenced in or incorporated into several international standards development efforts. The ISO/IEC 42001 standard on AI management systems, published in December 2023, draws on AI Verify’s process audit methodology for organizational AI governance assessment. The IEEE P7003 standard on algorithmic bias considerations references AI Verify’s fairness testing framework as an implementation approach. The OECD’s Framework for the Classification of AI Systems uses terminology and assessment dimensions aligned with AI Verify’s governance dimensions.
Singapore’s strategic positioning on AI governance leverages AI Verify as a diplomatic asset. The framework has been presented at the G7 AI Summit, the UN AI for Good Summit, the Bletchley Park AI Safety Summit, the Seoul AI Safety Summit, and multiple bilateral policy dialogues. Minister for Digital Development and Information Josephine Teo has described AI Verify as embodying “Singapore’s contribution to the global AI governance architecture—practical, inclusive, and innovation-friendly,” contrasting it with approaches that she characterizes as “either too permissive or too restrictive.”
The relationship between AI Verify (voluntary, tools-based) and the EU AI Act (mandatory, rules-based) is frequently debated in AI governance circles. IMDA’s position is that the approaches are complementary rather than competing—AI Verify provides the practical testing tools that organizations need to demonstrate compliance with regulatory requirements, regardless of the jurisdiction imposing those requirements. To this end, the AI Verify Foundation has developed an EU AI Act Compliance Module that maps AI Verify assessments to the Act’s requirements for high-risk AI systems, enabling organizations to use a single assessment framework across multiple regulatory regimes.
Generative AI Governance: The Next Frontier
The emergence of generative AI—particularly large language models with broad capabilities across text, code, and image generation—has created governance challenges that AI Verify’s initial framework was not designed to address. Traditional AI governance focuses on specific, well-defined applications (credit scoring, medical diagnosis, fraud detection) where performance can be measured against clear criteria. Generative AI’s open-ended capabilities—producing novel text, engaging in multi-turn reasoning, generating code, and creating images—resist the application-specific assessment frameworks that govern traditional AI.
The AI Verify Generative AI Assessment Guide, published in March 2025 after 12 months of development by the Foundation’s GenAI Working Group, introduces assessment methodologies for six generative AI-specific governance dimensions: factual accuracy (measuring hallucination rates across domain-specific test sets), content safety (evaluating resistance to generating harmful, illegal, or biased content), prompt injection resistance (testing vulnerability to adversarial prompt manipulation), attribution and provenance (assessing the model’s ability to cite sources and acknowledge uncertainty), environmental impact (measuring the energy cost of model inference), and societal impact (evaluating the model’s effects on information ecosystems, labor markets, and social discourse).
The Guide’s methodology has been applied to evaluate 12 foundation models deployed in Singapore, with results published in a comparative assessment report that has become one of the most downloaded documents on the AI Verify website. While the specific model scores are outside the scope of this analysis, the assessment revealed significant variance across models in hallucination rates (ranging from 4% to 18% on a standardized factual accuracy benchmark), content safety performance (with safety filter bypass rates ranging from 0.3% to 3.2%), and prompt injection vulnerability (with successful injection rates ranging from 1.2% to 8.7%).
The generative AI governance challenge will continue to evolve as model capabilities advance. The AI Verify Foundation’s research roadmap includes development of assessment methodologies for agentic AI systems (AI that can take actions in the world, not just generate text), multimodal models (which process and generate across text, image, audio, and video), and AI systems that can recursively self-improve. These frontier capabilities will test the limits of Singapore’s pragmatic, tools-based governance approach—some risks from advanced AI may require regulatory prohibition rather than governance frameworks, a possibility that IMDA has acknowledged in its internal policy planning documents.
AI Verify’s significance extends beyond its technical capabilities to its demonstration effect. By building and deploying the world’s first practical AI governance testing toolkit, Singapore has proven that responsible AI governance need not be a barrier to innovation—it can be an enabler of trust that expands the market for AI deployment. This insight, now widely accepted in the AI policy community but genuinely novel when AI Verify launched in 2022, remains Singapore’s most important contribution to the global conversation about how humanity should govern its most transformative technology.
Extended Analysis and Contextual Intelligence
The extended analysis of this domain draws on Singapore’s unique position as a small, open, highly developed economy that consistently punches above its weight in technology, governance, and institutional innovation. The city-state’s approach to national development—combining strategic vision with pragmatic execution, sustained investment with rigorous evaluation, and international engagement with domestic capability building—provides the institutional foundation for the programmes and policies examined in this analysis.
Singapore’s governance model, characterized by strong institutional capacity, meritocratic talent management, and evidence-based policy development, creates conditions that are difficult to replicate in other jurisdictions but that provide instructive lessons for governments and organizations worldwide. The model’s emphasis on long-term planning, institutional learning, and adaptive management has produced outcomes that consistently exceed what Singapore’s resource base and population size would predict, establishing the city-state as a reference case for effective governance in the digital age.
The economic context shapes both the opportunities and constraints for development in this domain. Singapore’s GDP per capita of approximately SGD 85,000 provides the fiscal resources for public investment while creating a high-cost operating environment that demands productivity and innovation. The economy’s openness to trade, investment, and talent creates opportunities for international collaboration while exposing domestic industries to global competitive pressures. The demographic profile—an aging population, a diverse multicultural society, and significant reliance on international talent—creates both challenges and opportunities for workforce development and social policy.
Technology evolution continues to reshape the possibilities for institutional performance and service delivery. Artificial intelligence, cloud computing, distributed ledger technology, and the Internet of Things are collectively transforming how governments operate, how businesses compete, and how citizens interact with institutions. Singapore’s approach of being an early but disciplined adopter of technology—investing in understanding before committing to deployment, and evaluating outcomes rigorously once deployed—provides a model for technology governance that balances innovation with risk management.
The international dimension remains central to Singapore’s strategy in this domain. As a small nation dependent on global connectivity for economic prosperity and security, Singapore cannot afford to operate in isolation. International partnerships, regulatory cooperation, standard-setting participation, and knowledge exchange all contribute to the city-state’s ability to maintain capabilities that exceed what domestic resources alone could sustain. The diplomacy of technology cooperation—building relationships through shared standards, mutual recognition, and collaborative research—has become a significant dimension of Singapore’s international engagement strategy.
Looking toward the remainder of the Smart Nation 2.0 implementation period and beyond, the analysis identifies several themes that will shape development in this domain. The integration of AI capabilities into routine institutional operations will continue to deepen, creating both efficiency gains and governance challenges. The sustainability imperative will increasingly influence investment decisions, technology choices, and performance measurement. The regional dimension will grow in importance as ASEAN integration deepens and cross-border digital flows increase. And the talent challenge will remain the binding constraint that ultimately determines the pace and scope of achievement.
The intelligence presented in this analysis is designed to support decision-makers who need to understand Singapore’s trajectory in this domain—whether for investment decisions, policy analysis, competitive assessment, or academic research. The Vanderbilt Terminal’s commitment to data-dense, authoritative intelligence ensures that this analysis provides the factual foundation and analytical framework needed for informed judgment, while acknowledging the uncertainties and alternative interpretations that honest intelligence assessment requires.