The Three Cs of Innovation: Caution, Curiosity, and Confidence in New Zealand’s AI Approach

Andy Neale, Deputy Chief Executive of Access and Digital Services at the Parliamentary Counsel Office, discusses how the right pace of innovation is guiding their approach to drafting and reform

When it comes to adopting AI in legislative drafting, Andy Neale believes in staying within what he calls the “Goldilocks zone” – “not too fast and not too slow.” As a Deputy Chief at New Zealand’s Parliamentary Counsel Office, Neale has been leading a series of AI experiments grounded in three guiding principles: caution, curiosity, and confidence.

Andy Neale is the Deputy Chief Executive of Access and Digital Services in New Zealand’s Parliamentary Counsel Office.

New Zealand has a unicameral parliament, and the Parliamentary Counsel Office plays a central role in turning policy into law within this streamlined structure. “We’re not AI experts. We draft the law,” Neale told me. “We need to learn more about this technology and understand it ourselves.” That curiosity has led his team to conduct five structured experiments with large language models – from using AI to check plain language compliance to generating explanatory notes and visualizing legislative amendments. But caution remains central. “We’ve learned not to give AI access to powers it doesn’t need – that can cause unintended consequences.” For Neale, the goal is not to replace people but to explore how AI can assist in managing growing workloads: “If a tool reduces the time needed to write explanatory notes by half, that’s a win.”

In this conversation, Neale offers a grounded perspective on what it means to apply new technology to legislative processes – and why staying curious and deliberate and confident about what they can bring to this challenge is the best approach forward.

---

Beatriz: The first question is for readers who are probably not familiar: could you explain the role of the Parliamentary Counsel Office in New Zealand's legislative process?

Andy: The primary role is to produce legislation for our government – whether for example that’s government agencies or ministers. When there’s a particular policy that needs to be developed into law – whether it’s a new law, a change in law, or removing a law – we’re instructed to draft that legislation, to draft the bill.

We then take it through the process of getting it ready to be introduced into Parliament for the parliamentarians to decide what to do. We also have functions that publish and make the law accessible. We manage the New Zealand Legislation website, and we make legislative data available to different users through different channels.

Beatriz: So you work very closely with Members of Parliament and ministries?

Andy: And with government agencies themselves. Most often, the instructions to draft will come from the relevant Ministry or agency.

Beatriz: How has the office's work evolved in recent years with the availability of legislative data and digital tools?

Andy: We were early adopters of legislative drafting tools and technology. Going back to 2008, we transitioned to a digital environment. We adopted XML as the basis for drafting bills, whereas many offices around the world will use Word documents or word-processing tools. I think we were one of the early adopters of XML for legislation. XML is a data format. You’ll have heard of HTML – the structure of the web. XML is similar, but it’s a way of capturing the structure of data or a document. That’s why we can talk about legislation as data. It’s not just text on a page. Underneath that page, the structure is organized. For example, the title is in a special tag that identifies it as a title. That structured format lays a useful foundation for this next generation of AI technology, which seems to work better with structured data.

Beatriz: Why do you think the office adopted this digital environment so early? Is that a characteristic of the office – always seeking innovation? I’m asking because I want to understand what prompted you to start exploring the use of AI in drafting. What was the turning point?

Andy: In general, I suspect our government is not dissimilar to others in that there's always been opportunities to take advantage of technology. Going back a long way to the days of the typing pool – where we shifted from handwriting to typed text – that was just following the trajectory of government and society in general. I don’t know if there was anything special about it. If anything, you'd say the conventional wisdom is that government usually lags behind the private sector.

In more recent history, regarding opportunities for AI, we’ve witnessed the rise of generative AI technologies. That's what really signaled a shift for us. These technologies are based on large language models. And law has language at its core, right? The previous era of AI was more traditionally number-based, which was perhaps less interesting to us. But the moment we saw technology that had language at its core – that changed everything. The legal profession is widely identified as one that’s open to early disruption because of its basis in language. We immediately saw the opportunities – and challenges – that came with that.

I wouldn’t say we’re an early adopter within the New Zealand government space, but I’ve pitched that we try to stay in the “Goldilocks zone” – not too fast and not too slow. We’re not set up for bleeding-edge innovation. That’s not our mission. But going too slow and just waiting puts us at a disadvantage. We try to stay curious, active, and see what we can progress within our remit without sitting around waiting for someone to tell us what to do.

Beatriz: That’s a great way to put it – the Goldilocks zone. And honestly, it's the first time I’ve heard someone directly link language models to legislative institutions and their core work, which is language. You shared on LinkedIn two recent experiments with AI. Could you walk me through what you tested, what you found, and whether it was useful?

Andy: First, I’d say our work with AI is very much framed as research and development (R&D) at this stage. I often tell the office: we don’t know what we don’t know. We’re not AI experts. We draft the law. We need to learn more about this technology and understand it ourselves. We don’t want to be 100% reliant on external experts to come in and tell us what to do. We need to be able to understand what’s going on and make our own strategic choices.

A big part of R&D is to support our own knowledge building. The actual technology may or may not be useful and that’s okay. We’re not necessarily looking to immediately turn a prototype into a production service. It's okay to learn something and discard the technology if it helps us get where we want to go. That said, we might strike gold. We might come across something that works brilliantly, and we can take further steps.

Beatriz: Just to clarify, are you using AI in your daily routine now?

Andy: As part of the daily routine, no, not yet. Our staff have access to some of the online tools, like Copilot, but they can’t use them with sensitive information. In terms of actual drafting, there are a lot of security and sensitivity issues. Aside from general access to generative AI, we’re not using it in our core business yet.

Beatriz: Got it. But you said you’ve done five experiments?

Andy: Yes, and we’re releasing updates on them, one each week. We looked at our end-to-end workflow, from when an agency instructs us to draft a law through to drafting, editing, publishing, and making it available to the public. We picked points along that workflow where we thought there were big opportunities for using large language models.

The first experiment looked at plain language assistance. We have a goal to produce law in the clearest way possible, and we have plain language guidelines. We tested whether an AI assistant could review a piece of drafted text and check it against our guidelines – things like avoiding jargon or using active language. The tool reviewed the text and offered suggestions for improvement.

The second experiment looked at whether AI could generate explanatory notes. These notes accompany a bill to explain what each clause is trying to achieve. It’s a manual, time-consuming process. But generative AI seems good at summarizing and interpreting. So we asked: could it help here?

The third experiment asks if we can automatically generate a prospective consolidation from a piece of principal legislation and an amendment. That means showing what the law would look like after a change. Our drafters often get asked for prospective consolidations to visualize the impact of proposed changes. The tool acts like a change tracker, showing strikethrough and replacement text. It makes things a lot easier – and we don’t currently have tech that does change tracking this way.

We also have historic legislation going back to the 1800s, where not all consolidations were created. We’ll never have the resources to go back and do them all. But an automated tool could be useful for researchers or anyone trying to understand what the law was at a given point in time.

Beatriz: Can you give a sneak peek into experiments four and five?

Andy: The last two focused on access, not drafting. The fourth explored whether AI can better organize and categorize our legislation – a kind of behind-the-scenes optimization. The fifth tested whether we can place a chatbot in front of our legislation to provide a service where the public could ask questions. But here’s the challenge: we can’t allow any chatbot we operate to interpret the law or give legal advice. We’re a drafting office, so it’s not for us to offer judgement. A chatbot would need to stay focused on answering straight-forward questions about legislative content that is grounded and referenced directly in the text. To do that well, the legislative data needs to be structured differently. For example, as a knowledge graph which maps explicit relationships between different types of legislation and its provisions.

We’re exploring building a database system to represent legislation in this way. And with the chatbot, the focus was to consider guardrails to keep it safe – not giving legal interpretations but offering helpful, accurate information.

Beatriz: What have you learned from these five experiments overall?

Andy: We're still formally evaluating the results. We have prototypes and proof of concepts. Now, small groups of staff are testing each tool to see what works and what doesn’t. But preliminarily, we’ve learned that for these tools to be truly effective, we need new ways to organize legislation. We’ve historically had a document-centric approach – acts, bills, and amendments as documents. But AI systems work better when things are chunked into smaller pieces. We’re exploring a fragment data model alongside the traditional document model.

We’ve also learned that AI does a decent job in the first experiments – like with plain language or explanatory notes – but then it needs work to handle complexity. While the R&D has been valuable, there’s more to do. We hope to pick one or two areas where we see the best return and invest effort there.

Another key learning is around control. We started with the principle that legislative technology is critical civic technology – core to democracy. It’s important that we maintain full oversight. We’ve looked at using open-source models (like those from Meta among many others) that we can install and operate in our own hosting environment. That’s been encouraging. It means we can leverage international efforts while maintaining local control.

We’ve also noticed that models are improving constantly. Some of the problems we can’t solve today might be solvable six months from now. For example, if a task currently requires a large, expensive model, in a few months a smaller, more efficient one might suffice. A strategy that is interesting to us is where we might integrate the tools into workflows and regularly upgrade the models.

Beatriz: It seems you view AI as a complement to your work, not a replacement.

Andy: Yes, exactly. It’s early days, and no one knows the full impact. But for now, a lot of human judgment is still required in developing legislation. We see AI as a tool to help our staff stay on top of growing workloads and build on our quality practices. The amount of legislation required each year is increasing, but we can't hire infinitely. As mentioned earlier, if a tool reduces the time needed to do something by half, that could be a win.

Beatriz: If and when you decide to adopt one of these prototypes, what safeguards would you put in place?

Andy: Several. First, security and assurance. Any new technology must go through reviews, and we’d likely isolate it so it only accesses what it needs. We’ve learned not to give AI access to powers it doesn’t need, that can cause unintended consequences. We’ll also need privacy reviews, though most of our work isn’t personal data – it’s public law. But we still need to assess privacy risks.

And importantly, any system would need human oversight. One more point: the effect on people’s capabilities. There’s research showing that offloading tasks to tech can reduce cognitive ability. Think of calculators and doing math by hand. We want to ensure that there are not unintended effects on the ability to retain human capability. We don’t know the answer yet, but it’s something we need to think through.

Beatriz: That’s very interesting. I was actually talking to a journalist today who asked whether legislators investing all of their time in social media content is a distortion of parliamentary function. My answer was similar: for new members, that’s the only parliament they’ve known. The same applies here: for people entering parliamentary services now, the tech shapes their understanding of the work. That changes how we approach training and strategy.

Last question: are you collaborating with other parliaments doing similar work?

Andy: We've been very inspired by others. For example, the UK’s work on explanatory notes. I also visited the European Commission a while back. They were generous in sharing their ideas. We have more regular contact with Singapore and Australia, especially through the Commonwealth network. We're not yet actively collaborating with anyone, but we’d love to. That's why we publish not just our research reports but also source code, so others can replicate or build on our experiments. I recently spoke with other Australasian colleagues about this and it’s a topic of ongoing conversation.

Beatriz: That sounds like a very productive way to collaborate. Is there anything I didn’t ask that you think is worth highlighting?

Andy: Just to reinforce that this is all still exploratory for us. We’re using the experiments to build up our understanding and to identify requirements for future systems. We could’ve hired a business analyst to define those requirements. Instead, we’ve used practical experiments, which I think is a very effective way to figure out what we actually need next.

Beatriz: That’s interesting because, as a political scientist, I really like your approach. It’s a combination of caution and curiosity, which I also adopt in my own work.