AI Verify Singapore Explained: How the AI Testing Framework Works

AI Verify is Singapore's AI governance testing framework and open-source toolkit, built by IMDA and now developed through the AI Verify Foundation, a not-for-profit subsidiary of IMDA. It checks an AI system against 11 governance principles using process checks, evidence and technical tests, while Project Moonshot tests applications built on large language models. It is free to use, and it is not a certification.

At a glance

How AI Verify came about

AI Verify began as a governance tool for traditional AI and has grown into a small ecosystem. According to the framework page and the foundation's programme pages:

• 25 May 2022: IMDA and the PDPC launched AI Verify as a minimum viable product for traditional AI.
• 7 June 2023: the toolkit was open-sourced on GitHub, and it is now developed under the AI Verify Foundation.
• February to May 2025: the Global AI Assurance Pilot paired AI assurance and testing providers with organisations deploying generative AI; 17 pairs were showcased at Asia Tech x Singapore 2025.
• 29 May 2025: the testing framework was updated to cover generative AI as well as traditional AI.
• Now: the pilot has become the Global AI Assurance Sandbox, and the foundation runs an AI Tester Accreditation Programme for testing firms.

The 11 principles and what gets checked

The framework has four elements: principles, outcomes, processes and evidence. For each principle there are desired outcomes, process checks that show you are working towards them, and documentary evidence that the processes exist. Where relevant, technical tests add measured results. The framework page lists these process checks per principle:

A process check is a concrete question with evidence behind it: is the system's intended use documented, can a person override its decisions, is there a record of the data used to train it. Most of the work is gathering documents you should already have, which is why a first pass often exposes gaps rather than test failures.

The foundation says the framework aligns with frameworks from ASEAN, the EU, the OECD and the US, and it publishes crosswalks to the NIST AI Risk Management Framework, the NIST Generative AI Profile, the Hiroshima Process International Code of Conduct and ISO/IEC 42001:2023. If you are building an AI management system or answering customers who use those frameworks, the same evidence does double duty. Our guide to the NIST AI Risk Management Framework explains that framework in plain English.

Testing traditional machine learning with the toolkit

The AI Verify Testing Toolkit runs technical tests for fairness, explainability and robustness on traditional machine learning models, records the governance process checks you complete, and produces reports you can customise. Its tests are black-box: they look at what goes in and what comes out rather than at the model's internal code. The toolkit is open source under the Apache 2.0 licence, with its code on GitHub, and it can be extended with your own test algorithms and checklists.

A typical use is a model that makes or supports decisions about people, such as shortlisting job applicants or flagging insurance claims:

• Fairness tests check whether outcomes differ across groups you define.
• Robustness tests check how much predictions shift when inputs are slightly disturbed.
• Explainability shows which input features drive the model's decisions.

Testing generative AI apps with Moonshot and the Starter Kit

For chatbots, copilots and other apps built on large language models, the relevant tools are IMDA's Starter Kit and Project Moonshot.

The Starter Kit is a set of voluntary guidelines for testing LLM-based applications before deployment. It focuses on five risks:

1. Hallucination and inaccuracy
2. Bias in decision-making
3. Undesirable content
4. Data leakage
5. Vulnerability to adversarial prompts

Its approach has three steps: identify the risks that matter for your app and set thresholds, test in a structured way from the app's outputs down to its components, and assess whether the results meet your thresholds.

Project Moonshot is the open-source toolkit that puts this into practice, with benchmark testing and red teaming. It implements the benchmarks the Starter Kit recommends, and the foundation pairs it with an AI Verify process checks tool for the governance side.

One lesson from the 2025 pilot is worth copying: test the real application, not just the underlying model. Your retrieval sources, system prompt, guardrails and user interface all change behaviour. Alongside the standard benchmarks, build a small test set from your own domain: real customer questions with expected answers, questions the app should refuse, and attempts to make it reveal confidential data.

Who uses it, and does a small team need it?

AI Verify is a tool organisations choose to use. Anyone can use the framework and toolkit, and the foundation's FAQ lists system owners, technology providers, researchers and service providers among them. Three points to keep straight:

• It is not a certification. The accreditation programme recognises testing firms, not AI systems, and the foundation states that accreditation is not a certification that tested AI systems are completely safe or free from all risk.
• Results are evidence, not a pass mark. You decide your thresholds and remain responsible for accepting the remaining risk.
• It complements the rules that bind you. It does not replace the PDPA or your sector regulator's requirements. Our explainer on Singapore's generative AI governance framework shows where testing fits in the wider picture.

AI Verify is most worth the effort when:

• you sell an AI product to enterprises or public-sector buyers who ask for evidence of responsible AI;
• your system makes or supports decisions about people, such as hiring, credit or claims;
• you already work to ISO/IEC 42001 or the NIST AI RMF and want test evidence that maps to them.

A small team using an off-the-shelf chatbot does not need the full toolkit. It does benefit from the process checks and the Starter Kit's risk list, which take hours, not weeks.

For higher-stakes deployments, the AI Tester Accreditation Programme helps you choose a third-party tester. The foundation says accreditation is valid for 12 months and that it charges no application or accreditation fees. The sandbox, which pairs deployers with testers and publishes case studies, expects testing and the report to be completed within about three months.

Practical first steps

1. List your AI systems and pick the one with the biggest effect on customers or staff.
2. Read the 11 principles and mark which apply to that system.
3. Answer the process checks honestly and note every gap.
4. For an LLM app, use the Starter Kit to choose risks and thresholds, then run Moonshot benchmarks and a red-teaming session.
5. For a traditional model, run the toolkit's fairness, robustness and explainability tests.
6. Write up the results, fix the gaps, and repeat after any change to the model, data or prompts.
7. For high-stakes systems, consider the sandbox or an accredited tester.

How CertTulen Academy can help

CertTulen Academy runs live, online, MCT-led training with hands-on labs. Technical staff who will test AI systems need to understand how models work first, and CT-AI-100: AI Foundations: Concepts and Architecture covers that over four days. Business and risk staff who review AI outputs can take CT-CLA-FLU2: AI Fluency: AI Capabilities and Limitations, a half-day course on where AI tools succeed and fail. We also run private cohorts for teams. Contact us for a batch that suits your time zone.