OTTAWA — The head of the British digital service is bullish about enhancing her government’s online relationships with the public, but deeply cautious about using artificial intelligence tools to do it.

Christine Bellamy became chief executive last June of the Government Digital Service—an inspiration for similar groups of insurgent tech advocates in multiple Canadian governments—after a stint in charge of the sprawling GOV.UK website and a decade in senior roles at the BBC.

She spoke to The Logic in Ottawa, where she was taking part in a conference on digitalizing governments, about changing culture and operations in institutions with centuries of history, and how to use generative AI in a setting where you can’t afford to make many mistakes.

This excerpt from the interview has been edited for length and clarity. 

How do you think about the challenge of integrating artificial intelligence and things like chatbots into digitally delivered government services?

Where I think we can get the most traction at the moment is very much more civil-service–facing. There’s less risk. It means, for example, when we write our content or we think about our policy, we’re able to think about the toil within that situation to see if we can move ourselves further through the journey.

We’re also doing some good work in the U.K. around how we can use AI to start to think about issues such as [detecting] fraud. Some of that is building on AI that’s been around for a while.

Recognizing patterns and so on.

Absolutely. Then you get to the user experience, and my view—this is my view, not representing the U.K. on this one—is, this is something we should be thinking and exploring about.

One of the biggest problems a user has when they come to digital services is they just don’t know where to start. So they start by, at the moment, probably typing a sentence into Google or whoever, and it takes them on a journey. If we’ve done our work, they should get there quickly.

But we’ve still got all of this noise that we’re not meeting people where they are and directing them quickly to what they need. That is an interesting space to start thinking about the application of AI.

If you’re a hairdresser and you’re running a small business in the U.K., you’d probably have to interact with 10 or 12 different parts of government. In this instance, the service has the opportunity to say, “Ah, OK, I’ve asked, I know who you are, I know a little bit about you, I can start to get you to where you need quicker, because I’ve got your context and I’m navigating you through it.”

It sounds like you’re talking about using AI tools to demystify the complexities of dealing with the government.

We have an opportunity to think about that in almost a walled-garden scenario. So although that’s still risky, it’s less risky than a user reading it on another AI platform, where our information is taken in with other people’s information, which might be misinformation.

We’re using our own content within GOV.UK, for example. That’s all content that is open and published already. Even that’s not straightforward, but the models, I think, will improve. They’ve improved so much in the last year. We need to work with the technology and see how it progresses.

We’ll work closely with other private-sector organizations and other governments to just think about how to do this work without doing anything that we think will be risky for users.

You’re dealing with a defined box of puzzle pieces. You know what they make, which is ideally a functioning government service, and you can use AI as an agent to assemble it for people.

I think our legacy of all these years behind us, all the things we’ve published, becomes an asset, not a hindrance, because we’re starting in the right place. The work we’ve done on things in the U.K., for example, on the design system, user-centric content, which we’ve led on, means that we’re also starting with content that has been written for the user to understand it. So we get the benefit of that next time around.

Obviously large language models will still shape it, but they’ll start with a better content store, in effect, which has already been signed off by many human beings. It’s already been through the wringer.

So generative AI feels like it’s worth experimenting with—and I would be the first to say, we will do experiments, and if they don’t work, we’ll stop.

What are your thoughts on using AI to make decisions?

I think where we are now, with people—experts—being involved in the processes, is vital.