The government is buying the shiny, expensive AI toys rather than the things that work, Tim McNamara says

This government's first year has been about saving money. We cancelled the ferries. We've cut health sector jobs. We've made thousands of officials redundant. In September though, the government announced the GovGPT chatbot. 

GovGPT is not free. It's an expensive toy pilot project that carries a lot of risk. To build your own AI service, you start by indexing all of the content and store it in a bespoke database. Then you work on training a bespoke model by tweaking a public model with your own data and teach it how to speak. To host a service, you need lots of testing to make sure it doesn't do something unintended, like tell people that they should die. This all effort and expense.

And what happens after all of that effort if no one ends up using it? You've just spent money on an shiny, expensive toy that gets kept on the shelf.

The easier way

There's a better way to leverage AI. Public sector agencies can publish accurate content. Plain old HTML works very well. Glossy PDFs work less well. You'll get the same results as doing everything yourself, at a fraction of the cost.

Putting content online allows all AIs to benefit, makes search engines more useful, and is extremely cheap to do. It pushes the cost of maintaining the model, hosting unfamiliar databases, and then supporting a live service, to the private sector.

Information rich

Agencies hold much more content than they realise. For example, every contact centre will have a script for answering common queries. There are other resources too. Training material will use language that's explanatory. Internal memos from senior management about changes to practice are useful context.

They could all be made public. Copy and paste will do, even if some of the markup doesn't carry over. The intended audience is machines that don't care a lot about formatting after all. 

This knowledge will then diffuse into the AI models within a few months at most. AI systems are becoming smarter too and are gaining the ability to actively search the web for new information.

Once the public sector's knowledge is public knowledge, Claude, ChatGPT and Copilot, Gemini, and every other model will be able to provide answers that are of equivalent quality to what the government could build for itself. Maybe even apps that people actually use, such as Facebook, Instagram and TikTok could provide answers to questions directly. Those answers could then provide links back to the original source - but only if the primary sources are available.

It will take courage and care to get more information released, particularly if there's less concern about presentation. Courage should be easier to find than money. Care is abundant in the public sector. The alternative to publishing accurate information is that models make things up.

Cash poor

Creating a bespoke model and hosting a bespoke service is extremely expensive. The companies producing these models have billions of dollars to invest. Callaghan Innovation does not.

There will be many happy vendors behind the scenes advocating strongly for the expensive option. It's expensive because vendors are on the other side of the transaction. The public sector spends far too much on failed ICT projects. It's foolish to think that this project will be different.

And it gets worse. GovGPT will be using a single base model, and therefore will be limited by its limitations. By staying neutral and focusing on being the source of truth, the government can benefit from improvements from any model, irrespective of the vendor/. 

Unlikely to be used anyway

The New Zealand public uses apps like Instagram, TikTok, and Facebook. They are not going to turn to some custom app. Attempting to create a single front door for government just creates another door.

The only way to get there in a scalable way is by keeping content online that's accurate, authoritative and accessible.

Boring technology works best

This knowledge is not new or exciting. About 25 years ago, the then Ministry of Commerce published web guidelines. This was then taken up by State Services Commission that championed what was then called "eGovernment". The fact that it's boring means that it's not fashionable. There's little money in commodity services.

By spending huge sums of money and lots of human capacity on something that's expensive and ineffective, the government is demonstrating its inexperience and incompetence in the area of artificial intelligence and digital technologies more broadly.

Tim McNamara runs the Accelerant.dev tech consultancy from Wellington. He's built software and data science systems for local businesses, global giants, and now his own startup. As an advocate for the Rust programming language, he hosts a YouTube channel, a podcast and gives talks and workshops worldwide. Connect with him on LinkedIn.