James Huang

Hello everyone, I'm James. Welcome to Digital Keywords , where we learn about tech — big and small. In recent years the letters "AI" have been in front of us almost every day. The full name is artificial intelligence — by definition, intelligence made by humans. But in this episode I want to play a pun that only really works in Mandarin: swap the character for "human/person" in "artificial" for the character for benevolence and care — "仁" — and you get "仁工智慧". One character's difference — what does it add? It adds a heart that cares for others. The same technology: you can choose to give it a caring spirit, or let it serve the rich without conscience. That wordplay is one of the things our guest has been most invested in these past few years.

Speaking of our guest Audrey Tang, many people may still picture mask maps and SMS contact tracing — pandemic tech. But Audrey is now a researcher at the Oxford Institute for Ethics in AI, running a programme called Civic AI — in Mandarin, 仁工智慧, a play on “artificial intelligence” with the character for care. That framework was formally launched at Oxford only this March; the book may not appear for another year or two. So this episode may well be Audrey's first in-depth conversation about this work-in-progress with a Mandarin-speaking audience. We're honoured to have her — welcome.

First I'd like to stretch the timeline for listeners. Before you joined the cabinet I had a chance to know you. Back then we were in g0v-related communities; I know you were in g0v very early and doing AI-related research.

So you were in touch with what we now call early AI very early. Later you became a minister without portfolio, then led the Ministry of Digital Affairs, until you stepped down and became our ambassador-at-large. Now you're researching at Oxford — could you walk us through that arc? What have you mainly been doing since leaving office?

Do you feel Taiwan is special — like a laboratory, or particular soil? Maybe because we got universal broadband earlier, and many people have various OSes, phones, computers, so we met digital tech — now AI — very easily, and hit similar social problems sooner?

It really is special. As you shared, Taiwan has often been relatively forward in adopting IT — we take things up quickly. We feel we're always looking up at the US and others — "they have this tech, why haven't we?" From computerisation to e-government and mobile, each wave — PC, internet, mobile, now AI — we've moved faster on that path.

But if so many countries, societies, and leaders want Taiwan's story, what path pulled you from today's AI toward people-centred Civic AI — when we say "仁", should we say mercy in English? What made you think we should do this?

So that's the thread. In much of our work together over the years — impacts on society — twenty-odd years ago people were already debating recommendation systems and related algorithms.

So the endless-scroll algorithms and later echo chambers — the Mandarin "同溫層效應" — are phenomena of the last decade or so, whether you call it big data, collecting signals, tuning parameters, training popular algorithms, especially when linking more people.

A striking phenomenon. But I notice something about your personality — my own read. When a system looks off here or there, most people blame or protest, demanding owners or developers fix it.

That's most of us. You've been different. Blaming is easier; you tend to think, discuss: rather than protest without acting, show a slightly different demo yourself. For example, early government platforms like the Join platform?

Do you see this trait in yourself? How do you keep turning, especially often ahead of the crowd, reminding us — shall we turn the picture, what shall we do? How does that thinking keep showing up?

That's serious cultivation. Many people face emotional swings and fast social change — especially AI's pace — we may all need to practise. One more thing: AI is moving very fast. For Business Next listeners, Andrej Karpathy — I mention him often — from vibe coding to agentic coding getting feasible and mature, plus harnesses and agentic workflows.

I wonder if you agree: compute is faster and faster, often far beyond most humans — for you maybe thousands or tens of thousands of times. Ways we used to judge right and wrong — pause, have slack — will human and AI speeds become more asymmetric? Do you already see problems, or early signs as AI shifted?

I'm still practising harnessing. Doing it ourselves, many people differ in grasping AI's pace or building a pipeline to ride it — AI agent, "little spirit", local lobster, whatever. Many will face that next.

How do you see superintelligence — AGI, ASI, superintelligence — industry pushing one single better stronger model as superintelligence? Or is superintelligence not the ultimate balanced direction ahead?

So it's an evolving story; now technology has emergent possibilities, and you're studying and proposing the framework we'll explain — from artificial intelligence to Civic AI.

From this framework to the Civic AI turn — how should we understand your framework and what Civic AI really asks us to care about, in Mandarin, English, and the world?

So you could say you validated it on your dad.

The kind of AI you just described is quite unusual, because most people looking at frontier models or frontier AI labs today are mostly commercial organisations. So there are a few traits: if you want to use it, you have to subscribe; the more you use it, the bigger the subscription.

And those companies — your subscription fee might not be all they want; they may want your data too — and as you said, with other business models possible, maybe ads, or recommending things to buy later, all sorts of methods. Or helping you complete tasks and then charging you through the outcome of those tasks.

We could talk about a lot of companies and different business strategies here, so I especially want to ask you — this is special — what you built, which your dad was originally using, or what's now popularly called this lobster kind of thing.

The Japanese sense of a local deity. This AI is relatively local, on-device. For this kind of local AI, are accessibility and privacy easy to set up, or will it still take a while? But more and more people should get to experience the privacy gap between big cloud AI models and a model on the ground that ordinary people can actually try?

That is something a lot of people will be discussing as they talk about AI — there are more choices, including the one you mentioned today. From different angles, especially on the technical side, we test things, so we roughly know what each model or technical stack can actually deliver. But I'd remind everyone to try more — there are lots of different tools, and open source offers many different options.

That leads to an interesting question. These last two or three years — since you stepped down — Taiwan's economic environment has shifted a lot; with semiconductors plus AI, the world needs all sorts of things from us, and that's one reason growth looks as strong as it does now.

But that growth runs into a lot of constraints. Why do they need us? They need compute — as we've been saying, we all hear about OpenAI, xAI and similar companies; in their development they all need a lot of compute.

But we're already seeing compute get stuck on many things — could limited power be one bottleneck?

And the next thing is we're finding it's not only power. When you build a data centre you need many things — networking, memory — so all these local facilities get constrained too; that's a bottleneck.

Now people like Elon Musk talk about xAI's next steps, including putting compute in space, even low-Earth orbit — all of that might happen. So there's a lot of discussion about bottlenecks: how to solve semiconductor production, methods like Terafab and various approaches — some stress TSMC or manufacturing capacity limits, some say hardware might ultimately be the ceiling on growth.

How do you see people discussing these limits — could it sometimes actually help draw a boundary for AI, and a boundary with human use — maybe that's the real line, not a bad thing, maybe a chance for everyone to hit the brakes?

That means we've watched too much sci-fi and too many sci-fi films.

So you think — like when we first encountered computers, the shift from mainframe to PC — it's a swing? People only think today's big models must be pushed with how much compute, or that training a model with however many parameters is what counts as good. But there's a lot of distillation down?

That's another huge shift. I'd especially like to invite you — in talking through Civic AI you used a very particular term, borrowing from ethics research: care ethics.

There's a philosopher in that field who's especially well known, Joan Tronto. Could you explain how you first introduced that into the Civic AI work you're doing now — how you understand her research when designing the foundations of Civic AI?

Right — usually either we've agreed on some declaration and you follow it, or you follow some framework; it's usually only that. She cares more about what you emphasise — care, or even relationship.

And we seem to grasp that kind of value very young.

But yes — if we ask others to try what Eastern filial piety means, that is harder. What does it mean?

While I was reading through your research, one thing I really admired was how you thought about taking these values and this framing to international audiences — not only within a Mandarin context, but across different cultural backgrounds and how we understand them.

I have to come back to something quite personal. In my own research I found that Taiwan has a lot of especially interesting value-driven design. Early on, when we talked about community care or community development, we often started from things like the digital divide. The easiest picture — and not hard to imagine — is remote areas where everyone gets fair access and can get online if they want to.

Right. I know you had to pay special attention to that when you were doing this policy work. But often we just pull the network out to every place, and the easiest way to let ordinary people actually use it is: I go there and build something for everyone to use. Later in my research I found that for social care, if you want people to use it, you go to the field where they already gather — maybe the library, maybe not. They do not only go to the township office; in Taiwan every place is special in that there is a faith centre wherever there is a settlement. Pull connectivity to the faith centre, pull it to the school — people usually show up there.

Those are the places where it is easiest to gather and to make things stick. Whether you are pushing digital policy, education, public health, long-term care, and so on — if you bring good resources to the right place, with appropriate design and guidance, you have a chance to make it work.

What I took from that later is a deep feeling: you do not need to build a so-called community centre from scratch. It may already exist — at the faith centre or nearby. Find the right point, work for that place, and you are more likely to be welcomed locally, accepted quickly, and to help people learn.

So I would like to ask you to introduce what you just mentioned — Civic AI from four care-ethics perspectives, from traditional care ethics to what we now call the 6-Pack of Care. Can we explain it in Taiwan’s local language: each of the six capacities, and how they connect to the good and bad of the AI everyone is using — so more people see how, in our relationships, we can apply values that already exist in society when we use AI, and what we might do or think differently?

Could you start by introducing the idea of the 6-Pack of Care?

Exactly — very split.

So four people, four parrots — or four little spirits — and they may each argue for a different direction.

Now many people argue on Meta, discuss on Threads, and call Meta AI out to look for consensus or possibilities.

So it is already in the data when you collect material at pre-training time.

In a sense, before AI, people might only gather at the temple gate at the village entrance. The loudest voice did not win; through discussion you found a few phrases or keywords of consensus, spread from there, and eventually everyone agreed on something. That is the most basic element of Attentiveness.

Next you mentioned Responsibility — what does that mean?