Context

Purpose

This guidance supports leaders, decision-makers and those in the New Zealand Public Service working or planning to work with Generative AI (GenAI). It enables agencies to explore and adopt GenAI systems in ways that are safe, transparent and responsible, and which effectively balance risks with potential benefits of these systems.

This guidance updates and expands on the ‘Interim Generative AI guidance for the public service’ published in July 2023 in response to feedback from agencies.

This guidance is part of the suite of Responsible AI Guidance for the Public Service which is being developed by the Government Chief Digital Officer to support the safe uptake of AI technology across government.

This suite of AI support material is being added to and enhanced over time and includes:

• a regular AI community of practice for digital practitioners in the public service
• an assurance regime for using and implementing AI in the public service
• a toolkit of practical and how-to resources on digital.govt.nz
• the Expert Advisory Panel for Public Service use of AI and the Public Service AI Assurance Regime and Framework (on its way).

This guidance has been developed alongside the Ministry of Business Innovation and Employment (MBIE) who are developing similar guidance for the business community. We’re also grateful for the collective expertise from across the public sector and their collaborative approach.

Support is available to help you and your agency develop knowledge and experience with using AI technology. We encourage you to connect and seek support from others on their AI journey and with us. For further enquiries or questions, contact us at the Government Chief Digital (GCDO) Office.

How to read this guidance

This guidance is structured into 2 sections:

• Section 1: GenAI foundations — key foundational aspects to underpin your customer experience and agency use. Topics include governance, security, procurement, skills, misinformation and accountability.
• Section 2: GenAI and the customer experience — key considerations that directly affect customer service experience with government. Topics include transparency, bias, accessibility and privacy.

You can navigate through the sections using the menu or the suggested related guidance at the end of each section. We recommend that you read the entire guidance to be aware of important factors that affect use of GenAI. Over time we’re building on the guidance, best practice, case studies and toolkits related to AI for the public service which is on Digital.govt.nz. Watch this space.

There are 2 key documents that should be read together with this guidance:

• Public Service AI Framework — which should underpin all AI work in the public service and is the basis of all practical support for public service agencies developed by the GCDO. It includes the OECD principles and it’s vision sets the direction for the safe and responsible use of AI to enable more efficient, responsive and high-performing public services.
• OECD AI system lifecycle — which should be referenced during ongoing development and use of GenAI systems in your agency. The OECD model grounds New Zealand AI development to international best practice. Its key themes are throughout this guidance.

New Zealand’s approach to adopting AI

AI in New Zealand has been in use in various guises for some time. In June 2024, Cabinet agreed that New Zealand’s approach to adopting AI will be in accordance with the OECD AI Principles. These principles promote innovative, trustworthy AI that respects human rights and democratic values. This guidance indicates the relevant principles in each section.

• Full Paper — Approach to work on AI — MBIE (PDF 155KB)
• AI Principles — OECD

The GCDO is working in tandem with several key agencies to ensure there’s consistency and to reduce duplication of effort and resource. In particular, the GCDO has teamed up with the Ministry of Business Innovation and Employment (MBIE) who has responsibility for AI for business and the Ministry of Statistics (Stats) who has responsibility for data.

Guidance A3 Summary

The accompanying A3 presents an overview of the OECD AI system lifecycle and key themes of this guidance. Such themes should be considered throughout this lifecycle and the ongoing development and use of GenAI systems in your agency.

A3 Summary: Responsible AI Guidance for the Public Service: GenAI (PDF 215KB)

Introduction

What is GenAI?

The OECD has defined the following:

• Artificial Intelligence (AI) — a machine-based system that can infer from inputs and generate outputs for explicit or implicit objectives. Different types of AI systems vary in their levels of autonomy and adaptiveness.
• Generative AI (GenAI) — a type of AI system that can create or generate new content such as text, images, video and music based off models and patterns detected in existing datasets.

Further definitions of key terms related to use of GenAI can be found in our Glossary of AI terms.

Open or Proprietary, Public or Enterprise AI

AI has common themes and issues, but different types of AI models and data handling methods need specific rules and precautions. Knowing these differences will help you make better decisions about using GenAI in your agency.

There are 2 key types of AI models:

• Open type AI — are those that have released their models for public scrutiny. These models can be studied, modified or built upon. Open models can help foster transparency, innovation and diversity.
• Proprietary AI — are AI models that are owned. The details of the models are kept confidential and can use licenses or other restrictions to maintain intellectual property while allowing (often commercial) use.

Open and proprietary models can be used for both public and enterprise AI systems:

• Public AI — these are AI systems that anyone can use, often through a website or portal, to answer questions or prompts. They gather information from various sources to create responses. Users do not control the system’s setup, data, or limits, and cannot easily check where the information comes from or how it’s used.
• Enterprise AI — these AI systems are created and used within a specific organisation to answer their particular questions or prompts. Organisations can closely manage how these systems are set up, what data they use, and their overall purpose.

This guidance will specify when recommendations relate to particular AI models or types of GenAI systems.

Public Service use of GenAI

New Zealand government agencies have identified a wide range of AI use cases. In many cases, these agencies have found benefits to applying GenAI systems to clearly defined problems.

Agencies are at different stages of AI implementation and using AI to improve productivity, efficiency, and customer experience. This progress aligns with key innovations by AI systems in the areas of text, video, and voice. A summary of where and when AI is being used in public service can be found in the following reports:

• 2024 cross-agency AI survey: highlights
• 2023 cross-agency AI survey: highlights

Opportunities for the Public Service to use GenAI include:

• Boosting productivity and efficiency — reducing in administrative burden through automating search and information retrieval, assisting in creating summaries or routine analysis, and reducing the demand on people. The result of this is that work speed is increased and teams can get on with more core work and delivery. For example:
  • Detection of unusual patterns in large datasets
  • Enhanced content search
  • Automated meeting summary and action lists.
• Streamlining modernisation of legacy systems — using GenAI to assist in interrogating existing codes, patterns and architectures. The result of this is that teams can map new digital systems while maintaining existing services. For example:
  • Digital applications can be efficiently developed using automate code
  • Digital and data systems can be tested efficiently.
• Enhancing customer experience of government services — reducing call or information request wait times, triaging issues so people get what they need more quickly, while reducing costs to government. GenAI can also support public servants by completing or automating routine tasks, helping staff to focus on frontline services. For example:
  • Chatbots to help the public self-navigate through complex information to find the answers they need
  • Automated note taking of service centre calls, which saves time so that calls can managed more effectively
  • Assisting with high volume tasks such as OIA requests so that these queries can be answered more quickly

While the opportunity for use of GenAI in the public service is substantial, there are several risks to the use of GenAI that, if they were to materialise, could seriously damage public trust and be potentially cause harm to communities. This guidance will help you to manage these risks.

Things to remember — It’s vital that agencies:

• Actively assess and manage the risks early and throughout the life of your GenAI systems
• Be open and transparent about how they’re using GenAI
• Support the rightsizing of risk management to enable safe progress
• Provide adequate training on understanding, using and managing GenAI systems to ensure safe and effective use.

Tips for successfully using GenAI

Involve stakeholders at each stage of the AI lifecycle

The public service must ensure any use of GenAI is safe and responsible. We recommend uptake that builds on strong foundations and involves stakeholders, partners and their communities at every stage of the AI lifecycle. To ensure we follow international best practice and to build on our collective knowledge, we look to the align with OECD principles.

Key stages of the OECD AI system lifecycle to follow are:

• Plan and design — understand the scope and set key objectives and direction for the implementation of GenAI. Consider your model, and think about key foundations and building in privacy, security and accessibility.
• Collect and process data — ensure good quality data. This includes sensitivity labels and ensuring accurate metadata. Consider if there’s any personal information and robustly assess potential risks and benefits.
• Build and use a model — identify and train your preferred model. Consider iterating the model and review the data and prompts used to improve accuracy.
• Verify and validate the outputs — test your model and evaluate its outputs to assess performance across a range of scenarios and prompts. Consider if it’s performing as intended or if further iteration is required.
• Deploy in a production environment — move into a production environment. Consider how GenAI works with other systems and how staff are trained to use it. Be transparent about your use and seek ongoing feedback.
• Operate and continue to monitor for both intended and unintended impacts — Continually assess GenAI use, including intended and unintended impacts. Consider if you need to go back to an earlier stage and iterate to produce trustworthy and useful results.

OECD AI system lifecycle — OECD (PDF 149KB)

Implement guardrails for the safe, responsible and transparent use of AI

Ultimately, agencies are responsible for their use of GenAI systems. This includes how AI is adopted and used, and for what purpose.

There are considerations for agencies across this guidance and that apply at every stage of the AI lifecycle.

• Maintain trust and confidence — consider how to safely, responsibly and transparently trial and adopt emerging technologies like GenAI.
• Build by design — incorporate key foundations into your earliest thinking around using GenAI. You’ll find information around key foundations including information management, privacy, security and accessibility in this guidance.
• Engage early — work with key stakeholders, partners, teams and end users to ensure their needs and concerns are met when using GenAI.
• Establish strong foundations — GenAI involves new ways of working, and may require further training, improved data, changes to information management practices, or different architectures or patterns to see the best outcomes.
• Select the right tool for the task — test and prototype to confirm it’s the right tool for the desired outcomes before any significant investment.
• Usual rules apply — consider also existing guidance and obligations around using a new tool like GenAI. This guidance provides links to more information to get you started.
• Keep humans in the loop — involve people at all stages of designing, training, testing, evaluating and moderating GenAI systems and their outputs to drive best outcomes.

Connect with other practitioners

Finally, do not do it alone. GenAI is in use across the public service. Connect with others using GenAI in similar ways to share approaches, patterns and solutions. This will help grow our collective knowledge on how to best use these systems. For help on how to achieve this, see Next steps for safe, responsible AI in government.

This guidance supports agencies and it’s part of a suite of support, advice and best practice the GCDO is building over time. A mix of in person, online and research driven tools and best practice will only grow alongside this guidance. While we support agencies to be responsible for their own use of GenAI systems, we support system-led and collective sharing and growth of quality tech and leadership to grow delivery, service and efficiency. Throughout these sections, we encourage you to always undertake rigorous evaluation of any system or process relating to AI. Agencies should feel supported to continue their AI journey. If you have questions, contact the GCDO team at gcdo@dia.govt.nz.