What Is AI Verify, and Who Is It For?

What exactly is AI Verify?#

AI Verify is a structured way to show that an AI system is built and run responsibly. Its testing framework is a checklist of outcomes and processes under 11 internationally recognised governance principles, and each process asks for evidence. Open-source tools sit alongside it for technical testing. The AI Verify Foundation publishes it on its website.

The Foundation describes itself as a global open-source community that brings together AI owners, solution providers, users and policymakers. Its About page states that it is a not-for-profit, wholly owned subsidiary of IMDA, and that it promotes best practices and standards for AI.

AI Verify sits within a wider set of Singapore resources for AI governance. If you are new to the topic, start with what AI governance is and how organisations begin.

A short history#

How does the AI Verify Testing Framework work?#

The framework has four elements: principles, outcomes, processes and evidence. Each principle has desired outcomes, each outcome has testing processes you carry out, and each process is validated by documentary evidence. For every process you record whether it is implemented, and you explain your answer, especially when you mark it "No" or "N/A".

Here is how the four elements fit together, using the first transparency outcome on the framework page as the example:

1. Principle. Transparency: an overarching consideration the AI application should be able to adhere to.
2. Outcome. Find suitable ways in the AI lifecycle to tell end users and other relevant parties what they need to know about the system, such as its purpose, limitations and risks, in an accessible way.
3. Process. Design an in-house policy on communication to consumers that sets out the principles for transparency.
4. Evidence. Internal documentation, such as the policy document itself.

The framework page says it is good practice to give reasons when you answer "No" or "N/A", and to note where the evidence sits when you answer "Yes". That habit is what turns a checklist into a record someone else can review.

The framework comes in three versions. The PDF gives an overview. The Excel version lets you track progress through the process checklist. A software tool, for generative AI only, helps you document your practices and generate a summary report showing your alignment with the framework.

What are the 11 principles, and how many checks does each have?#

The 11 principles range from transparency and safety to accountability and human oversight. The framework page lists a number of process checks under each one, so some principles carry far more detail than others. The table below shows the counts displayed on the page, which helps you estimate where most of your evidence-gathering effort will go.

The Foundation's FAQ describes fairness as the mitigation of unintended discrimination, which links closely to bias in data and outputs. The framework page says it is aligned with frameworks from ASEAN, the European Union, the OECD and the US, and mapped to the NIST AI Risk Management Framework, the NIST Generative AI Profile, the Hiroshima Process International Code of Conduct and ISO/IEC 42001. These counts can change when the framework is updated, so check the live page.

What tools come with AI Verify?#

Two open-source tools support technical testing. The AI Verify Testing Toolkit tests traditional AI models for fairness, explainability and robustness. Project Moonshot runs benchmark testing and red teaming for applications built on a large language model (LLM). Both are published by the AI Verify Foundation and complement the process checklist.

Project Moonshot is linked to IMDA's Starter Kit for Testing LLM-Based Applications for Safety and Reliability, which IMDA describes as voluntary guidelines. The Starter Kit focuses on five key risks: hallucination and inaccuracy, bias in decision making, undesirable content, data leakage and vulnerability to adversarial prompts. The FAQ says the toolkit supports both third-party independent testing and self-assessment.

Who is AI Verify for?#

The Foundation's FAQ states that anyone can use the framework and toolkit. The framework page names four main audiences: AI application owners and developers, internal compliance teams, external auditors, and policymakers and standards organisations. In practice, it suits any organisation that wants a recognised structure for showing how its AI system is governed.

The FAQ adds technology providers and researchers who want to contribute testing algorithms, service providers offering AI testing and advisory services, and companies that want to build on the open-source toolkit.

For a small organisation, the most practical starting point is the Excel process checklist. You do not need to run technical tests to begin: processes such as the transparency example above ask for policies and records that a manager can gather.

What might AI Verify look like in a Singapore workplace?#

Picture a small insurance agency in Tanjong Pagar that plans to put a generative AI assistant on its website to answer common questions about claims. Before launch, the operations manager wants a clear picture of what is in place and what is missing. The AI Verify process checklist gives her a ready-made structure for that review.

This is an illustrative scenario, not a real case. A sensible approach would be to:

• download the Excel version and work through it with the colleague who set up the assistant
• start with transparency: is there a written policy on what customers are told about the assistant, its purpose and its limits?
• under accountability, write down who approves changes and who can switch the assistant off
• mark processes that do not apply as "N/A" and give a short reason for each
• note where each piece of evidence is kept, so a reviewer can find it
• look at Project Moonshot or the Starter Kit for testing ideas on hallucination and data leakage

The result is a gap list the agency can act on. It is a record of its own practice, not a stamp of approval from anyone else.

Common mistakes#

The most common mistakes with AI Verify come from treating it as a badge rather than a working tool. Organisations tick boxes without evidence, skip the reasons for "N/A" answers, or assume a completed checklist means the system is safe. The framework only helps when the answers are honest, the evidence is real and the review is repeated.

• Ticking "Yes" without evidence. Each process asks for documentary evidence. If you cannot point to it, the honest answer is "No".
• Leaving "N/A" unexplained. The framework page calls giving reasons good practice. A reviewer cannot judge a blank.
• Treating it as a certification. The framework page presents AI Verify as a way to assess and show responsible practice. It does not replace your own judgement or your obligations under laws such as the Personal Data Protection Act (PDPA).
• Doing it once. The framework was updated for generative AI in 2025. Review your answers when the system, the data or the framework changes.
• Skipping human oversight. Human agency and oversight is one of the 11 principles. Decide who acts as the human in the loop and reviews outputs before the system goes live.

What to learn next#

AI Verify works best alongside a clear risk process and a written policy. Use the checklist to show what you have in place, use a risk assessment to decide what matters most, and use a policy to tell staff what they may and may not do. The guides below cover each step in plain English.

• Work through a step-by-step guide to assessing AI risk, which draws on IMDA's Starter Kit and the NIST framework.
• Compare AI Verify with the management system standard in what ISO/IEC 42001 is.
• Browse every guide in the AI governance category, or step back to the AI learning pathway.
• To build and test a simple AI app yourself, with responsible use built in, see the GenAI Builder Sprint course.

Sources#

1. What is AI Verify Testing Framework, AI Verify Foundation, accessed 8 November 2026
2. About us, AI Verify Foundation, accessed 8 November 2026
3. FAQ, AI Verify Foundation, accessed 8 November 2026
4. AI Verify Testing Toolkit, AI Verify Foundation, accessed 8 November 2026
5. Project Moonshot, AI Verify Foundation, accessed 8 November 2026
6. Starter Kit for Testing LLM-Based Applications for Safety and Reliability, Infocomm Media Development Authority (IMDA), accessed 8 November 2026

Frequently asked questions#

What is AI Verify?+

AI Verify is a testing framework and set of open-source tools from the AI Verify Foundation. It helps organisations assess and document the responsible implementation of an AI system against 11 internationally recognised AI governance principles, for both traditional and generative AI.

Who runs AI Verify?+

The AI Verify Foundation, a not-for-profit wholly owned subsidiary of the Infocomm Media Development Authority of Singapore (IMDA). IMDA and PDPC launched the framework and toolkit as a minimum viable product on 25 May 2022.

Who can use AI Verify?+

The Foundation's FAQ says anyone can use the framework and toolkit. The framework page names AI application owners and developers, internal compliance teams, external auditors, and policymakers and standards organisations as its audiences.

What are the 11 AI Verify principles?+

Transparency, explainability, reproducibility, safety, security, robustness, fairness, data governance, accountability, human agency and oversight, and inclusive growth, societal and environmental well-being.

Does AI Verify cover generative AI?+

Yes. An updated testing framework released on 29 May 2025 addresses risks from generative AI, and a software tool for documenting practices and generating a summary report is available for generative AI applications.

What is the difference between the AI Verify Testing Toolkit and Project Moonshot?+

The AI Verify Testing Toolkit tests traditional AI applications for fairness, explainability and robustness. Project Moonshot runs benchmark testing and red teaming for LLM-based applications, using the benchmarks recommended in IMDA's Starter Kit.

Is a completed AI Verify checklist a certification?+

The framework page presents AI Verify as a way to assess, document and show responsible AI practice, not as a certification. A completed checklist records what you have in place; it does not replace your own judgement or your legal obligations.