Hi, I’m Audrey Tang, I’m a digital minister. I’m a non-binary minister, not just in gender, but also parties and really everything. I take all the sides.

[laughter] I don’t think I’ve had that as an introduction before. So, thank you, and great to have you here with us.

I’ve been in Taiwan a few days now, actually, and I might just jump into something that I was doing on the first days that I was thinking about on my way here to you.

I went to some civil defense classes, and at one of them, they talked about the possibility of a cyber-attack when it came to the threat from China being much more likely than a missile attack, for example. And I was wondering, are you thinking about that? In what way, and what would you do to protect the people of Taiwan?

Yeah, even though last August, when our ministry first started, the same hour as our website went online, missiles flew over our heads following Speaker Pelosi’s visit. That was just that day, but every day, according to our National Security Council, there’s five million cyber-attack attempts on our network, and so, when you talk about possibility, no, this is reality.

Every day, we’re facing that kind of attack, and last August, in that single day, the adversary threw 23 times more resources compared to the previous peak just to keep the Ministry of National Defense, the Foreign Service, and the Presidential Office website down for a few hours just to disrupt it. So, the kind of resource they’re throwing into cyber attack really is limitless, and so that creates a very different resilience requirement for Taiwan compared to other countries.

So, how do you do that? Because I’m just thinking of some of the young people I met that were really thinking about preparing for the first time, okay? And they were trying to become more aware of what to look out for, what might be disinformation, and what might be a cyber-attack,

So, what do you do as Minister for Digital Affairs to try and protect Taiwan, and I suppose also raise awareness for those people?

Yeah, no, we launched, for example, the 111 SMS service, and we need people to tell each other that if you see a governmental SMS saying that your tax is due or something like that, parking tickets… And starting next year, if it doesn’t come from a short number like 111, that’s fake, that’s cyber-attack, it’s a phishing attack, don’t click on any of those links.

So, we’re switching to a different default. Previously, people assumed when they received a message or encountered somebody online, they assumed it’s a human, until they exhibit behavior that looks like a bot, but nowadays, with generative AI, with voice-cloning technologies, there’s very convincing personas that are actually bots, but it will take hours and hours and even weeks for them to exhibit some behavior that’s bot-like, so we need to flip the default.

Everything is assumed to be a bot, unless it comes from a previously face-to-face meeting, we exchange mobile phone numbers, so of course that’s not a bot, or it comes from those short numbers with verified checks, with provenance.

And is that short numbers, the 111 that you talk about, is that cyber-attack-proof?

Yes, so we designed 111 to run on the system called the T-Road, which is like the Estonian X-Road. This is something that’s specific to governmental agencies.

So, this information exchange maintains integrity by working outside of the public internet, so that’s one. We also work with the three leading telecom providers, currently there’s five, but soon three, so that any foreign telecom providers, or even local ones, proposing to be a telecom provider via 2G signal hacks and so on, none of them can fake 111. You will instead see, like, plus 870111 or whatever, but there’s no way to forge that short number.

Interesting. As we talk about disinformation, or indeed cyber-attacks, you have used the phrase ‘humor over rumor.’ Talk me through that.

Sure, so during the pandemic, for example, we noticed that when people are anxious, they’re much more likely to repeat polarizing messages. And in Taiwan, for example, we don’t have a political anti-vax faction.

Really? Nothing?

Nothing.

Nobody?

Nobody. And the reason why is that we shifted this to a friendly competition, to my vaccine is better than your vaccine. And so, there’s four, and finally five, different vaccine brands to choose from. If you interview people around that time, there are people who will insist only the vaccine they got was effective, and the other ones are like, you know, water or something.

And just to make a point, I have been inoculated with four different vaccines…

Is that a joke?

That’s not a joke.

So, you really are non-binary.

I really am non-binary, and indeed plural, right? Digital in Mandarin is called ‘shu wei,’ which also means plural. So, I’m also a plural minister.

Okay, a pluralist.

So, the point I’m making is that with some humor, we can always turn a zero-sum competition, something that only has one winner and everybody loses, into a friendly competition where it’s a positive sum, right? If people keep saying, oh, actually, my friends all got this vaccine brand, and according to the open data, which is updated daily, that people in this age bracket far prefer this vaccine, and so on, they all end up getting some vaccine.

And so, although there is still some polarization, it is within a kind of friendly competition frame.

Are you able to use that methodology in any way when it comes, and I’m gonna come back to hostilities with mainland China, for example?

People have different views. Some people want the status quo. There’s some that might push for formal independence, and a tiny minority, really, that would talk about reunification. But there is a fear there, I think, when I speak to some people about what might happen in the future, particularly, I think, when they look at Ukraine or they look at Israel-Gaza, that’s what they’ve mentioned to me.

I don’t know, how do you get people prepared? How do you speak to them about a potential threat?

Plurality means collaborative diversity. So, it’s not just about collaboration, it’s also about respecting diversity. And the people you just mentioned that have very different views on the PRC. On the other hand, there are also people within PRC that actually want democracy, want democratic networks, and so on. And so, by investing in those pluralist technologies, we also enhance the ability to self-organize and to not be censored or tempered, and so on, within even the most autocratic regimes.

And so, this kind of people-to-people solidarity toward democracy, this is what unites people together. So, this is not about a political party winning and taking all the power and things like that, but rather about a commitment to simply say, actually, democracy can advance mutual understanding, and technology can advance democracy.

And I understand this almost sounds alien now, right, to a polarized polity, but this is what we’re aiming toward, because we do see autocracies want to spread this meme that says democracy only leads to chaos, only leads to polarization, only lockdown, shutdown, top-down control works. And so, we heard a lot of that during 2020 around COVID.

But our main idea of pre-bunking it is just showing that actually democracy can foster mutual understanding and collaborative diversity, one case in point which are marriage equality, right? So nowadays, I think a great majority, almost 70% or even more now, support marriage equality. But that was not the case during the referenda. People were very polarized. And using this kind of plural mindset, we identified the main division point was actually not about wedding two individuals, but rather joining two families in kinship.

So finally, our ratification after two referenda was such that those two persons, they wed, but their families don’t form kinship, family connections. And so basically, just bylaws, but not in-laws, like father-in-law, mother-in-law. And then that, I wouldn’t say pleased both sides, but both sides can live with it. And on top of that consensus, people start growing to like each other more.

So, some of the concepts you’re putting out are really what I would expect from a behavioral psychologist. I know you have been named by Time one of the most influential people in AI in 2023. Is it some sort of intersection between those two that drives you?

Yes, definitely. So, before I joined the cabinet, I worked with the Oxford University Press and also with Apple on Siri technology. Siri is this AI chatbot that doesn’t hallucinate as much as ChatGPT, but I digress. Because that was the important thing for us. And so, what we have consistently learned during this time was that AI is only useful inasmuch as it assists community and people to uphold their agency. People don’t actually want a super intelligence that makes decisions for them, but it’s okay if people make better decisions together, assisted by a facilitating AI.

And so, that was our main contribution, actually, to the current generation of gen AI, is just to get hundreds or thousands of statistically representative people through an online survey or through face-to-face deliberation, and we just converge everybody’s expectation about AI, and then we show the top AI labs, like OpenAI, Anthropic, and so, they can tune their AI to respond to the people’s needs.

So, this is, I think, a variation on the wisdom of crowds, in a way. So, that theory that assumes the knowledge of a crowd results in better decision-making or innovation or problem-solving than that of a single individual.

Some might say, is that not groupthink? That a lot of people in the same sphere are going towards the same way of thinking. I mean, when we talk about AI, some of the issues that you will know better than anyone, it’s only as good as the data that is inputted. And a lot of people, maybe, that are motivated to take part in these experiments are of a certain type, that they’re from a certain group.

When I said statistically representative, I literally mean sortition, like random numbers, like drawing lots. So, for example, both OpenAI and Anthropic, working with the Collective Intelligence Project, which I’m a part of, just did this survey within their user base or within the American public, but they ensure that there is statistically the same constitution of that 1,000-people group as the average American citizenry.

And that is the important part, because exactly as you said, if this is about who can spend most time on an online forum, you will get a very different self-selected group.

Okay, and let’s talk a little bit about your initiatives that you’ve had. I read about your alignment assemblies. Could you explain them to our global listener that does not know anything about them?

This is a very simple idea. AI, in its current form, you can tell it to behave in such and such a way. And it might remember for a chat session, but it will not fundamentally change its bias, because exactly as you said, it’s constrained by the kind of data it reads when it’s doing the pre-training.

On the other hand, the AI can be amended by what we call an adapter. With the same AI, like Anthropics Cloud, or GPT-4, or LLaMA, a local community can train an adapter that says, before this AI actually speaks to me, it needs to run it through this cultural perspective, this cultural lens, that takes these important things into account before actually outputting any message to me.

And to train this LoRA, we can do this adapter’s training overnight on a MacBook. So, on my laptop, I have trained an adapter based on my email correspondence that writes my email reply drafts for me. I always read before hitting send. It’s not an answering machine. But it really imitates my style, and my thinking, and so on, very, very well.

And so, because everyone can do this personal fine-tuning, alignment assembly is about getting a group of people, maybe in a community, maybe in a company, and so on, who have interacted with AI and noticed the bias, or the harm that could be caused by a certain worldview, or whatever, and then just vent, right? Just say what you don’t like about it, and what you like about it.

But at the end of the conversation, we would just take everything that was deliberated, and tune an AI based on the collective expectation of such communities. And this can be done not just on AI ethics, or whatever. There are experiments that was done based on the interviews of certain communities that are disadvantaged, or are suffering from systemic injustice, and so on.

And again, they can create an AI language model that’s like a speaker for them, that decision makers can query, can have an interactive debate, even. And they will get some sort of representation based on the synthetic avatar that is co-created by this community conversation.

So, it’s… I’m having a look at some of the examples there. Heal Michigan was one, another about incarcerated adults. So, it’s almost becoming, in this case, AI voice for grassroots movements?

Yes, exactly. So, if you’re a community facilitator, this is not replacing your work.

This is about, you probably already have hours and hours of deep interviews, or community forums, or deliberative workshops on YouTube or Vimeo, or somewhere. And then you feed the URL, the playlist, to this AI.

Then it’s just understand the perspective, and summarize it in a way that is sympathetic to that group. And you can then lead a group discussion, and show them the summary. And they can say, ‘oh, this was misrepresenting,’ ‘I didn’t mean it like that,’ and so on. And you can also tune it overnight. And then after a while, you will have a co-facilitator that understand very deeply what is important to the community for them to have a voice, exactly.

Okay. Let me talk about… You were a protester once.

Sure.

And now you are a government minister. There were many actually, from the Sunflower Movement that also did move into government roles.

How was that transition? Also, how does it feel, I suppose, to be in this seat after being in the movement previously?

Yeah, I always said that I’m a demonstrator, not a protester.

What’s the difference?

Oh, a demonstrator is somebody who demos, right? So, when we occupied the parliament back in 2014 for three weeks, totally nonviolently, the idea was not that we want to block the liberation on the cross-strait service and trade agreements. Rather, it is about a demonstration of a new way to deliberate about that trade agreement with the PRC regime.

And namely, that the 20 NGOs that gathered around the occupied parliament each held, a little bit like alignment assemblies, talking about specific aspects of the CSSTA. So, people would talk about labor conditions, they would talk about cybersecurity, of over-reliance of the then new 4G core network infrastructure, and so on and so forth. And once they reached some consensus, it was broadcasted, captured, documented, so that the next day’s conversation can begin with the consensus of the previous day. So over three weeks, we actually converged on a set of very coherent demands, which was then ratified by the head of the parliament.

And so, it was a demonstration in deliberation rather than a protest that blocks parliamentary progress.

So, you don’t like the word ‘protest’?

No, I like the word ‘protest’. It was just that I’m more about demoing a kind of process.

It is interesting, though, those aspects that you mentioned there as a demonstrator. Because I was reading a story actually today about diaphobia. It was a new word for me, and it’s people having a fear of real dialogue. And there’s various articles that are being written about it.

But I know you have looked at carefully as a minister for digital affairs about how people interact and how you can get over some of the disagreements, the polarizing issues that often happen at the beginning of a conversation. Do you want to expand on that?

Definitely. So, the way that I facilitate and design facilitation is inspired by the focused conversation method, which means that we establish the facts, the objective part, before we even talk about feelings. But once people are informed of the shared facts, we just talk about feelings. So, as I mentioned, it was venting, right? People need catharsis before going into ideation or any suggestion of possible solutions. And then finally at a solution-making stage, we look at the ideas and combine those ideas.

So, the important thing is that we don’t confuse personal feelings or opinions with actionable results. We don’t actually jump to solution. We want to make sure that people are in each other’s shoes, so to speak, so that people who are anti-marriage equality and for marriage equality, they may both believe a lot in family values. They just go into it in different phrasing, right? But once we have this safe space where people can express their feelings, they actually discover their feelings resonate with one another, and then we can talk about more nuanced or eclectic solutions that people can both live with.

It sounds very positive and optimistic. I believe you’re an optimist.

I am, of course.

But there must be certain issues that Taiwan is facing that you will never come to a resolution no matter what process you put in place through digital means.

Well, yes and no. I mean, the main thing that we solve through this method are what’s called overlapping consensus, meaning that instead of debating about the abstracted economic theory of sharing economy versus gig economy versus extractive economy, whatever, we just talk about very simple thing like, oh, there’s in 2015, UberX driver that don’t have a professional license that picks people on the street randomly via an app and charges them for it. What do you feel, right?

And it turns out people feel the same once you have this very specific case. People say don’t undercut existing meters. People say insurance, registration, and so on. But if you start a conversation saying, oh, what do you think about capitalism? I don’t think it goes anywhere.

Okay, so you feel it’s really about framing the question as well.

You do have the questions and people are speaking out. Freedom of speech, freedom of expression, freedom of assembly here in Taiwan. How would you describe the last few years for Taiwan’s democracy?

Well, yeah, I think in terms of, especially internet freedom, association, expression, and so on, Taiwan is currently ranked, according to Freedom House, the top in Asia. And it was not the case, actually, before pandemic. And…

What happened?

No, I think what we have shown was that while many countries were forced by the amount of what WHO calls infodemic, which is disinformation related to the pandemic, so much so that they have to impose some sort of control on the freedom of expression online, Taiwan actually relied on the freedom of expression to create, as I mentioned, the humorous friendly competition between vaccine types and so on, so that it actually advanced freedom of expression online.

So, while in our region, many countries backslid on their amount of free expressions, Taiwan actually advanced, not by a lot, I would say, advanced a little bit, but because there’s a gradual backslide across our region. So now we’re at the top one.

I looked at that list, actually, and you were ahead of the United Kingdom. I think I’ll need… Who was ahead of you? Iceland, Costa Rica, Canada? Have I got that right? I’m not sure about my number three. And number four, I think, was Taiwan. And that’s for Freedom House 2022 for internet freedom.

There was, with some of your initiatives for a digital democracy in Taiwan, a freedom of expression we’ve talked about, but no reply button on some of the projects. Talk to me about that.

Oh, yeah. Okay. So, we didn’t invent this idea in Taiwan. We adapted from Iceland, from Better Reykjavik. So, in the Better Reykjavik petition platform online, when you post a petition, people can second it with the supporting arguments, or they can say, no, we have other ideas. But in the display, they display these two in two different columns, and there’s no way to reply across columns. So, you get the best pro argument and the best con argument, but there’s no way to distract people from these arguments just by replying and saying, I don’t agree with you, and so on.

And this is really good in the sense that if you don’t agree with the best pro argument, you actually have to think what is the contra argument that will bring some of those supporters toward you. So, this is a way to incentivize bridge making, not polarizing.

A very similar idea, after we adopted this in our petition platform and our poll list platform, Twitter, now called X.com, adopted this idea into community notes. So, if you want to look at plurality technology currently reaching the most people, it’s probably community notes. So, if you have a tweet, and it’s just part of the story, or it’s horribly wrong, you sometimes see a note attached to it. But it is contributed by a fellow Twitter user that managed to convince people on both aisles. So, there is an open algorithm that measures the bipartisanship, the bridge making potential of a note, and they only feature those notes as community notes.

And this is very powerful, because those cruxes, the bridge makers, once they get encouraged, they will devote more time to bridge that difference instead of just replying on a reply thread and keep dunking on each other’s pro or contra arguments.

Why are you such an optimist, do you think?

Well, because we did counter the pandemic without a day of lockdown, and the infodemic without a single administrative takedown. So, we do have some confidence that this thing actually works in a polity of 23 million.

What is the biggest challenge?

The main challenge is to maintain credible neutrality. So, I think now all the four major parties, MPs in the parliament, generally understand that I’m not just non-binary, but really all-partisan, right? Non-partisan, pan-partisan, so that I take the sides of all four major parties.

But back in 2014, it was considered almost impossible for someone to have credible neutrality and also work on internet style of platforms and digital democracy.

And there is an election looming in January. If the party changes, then what happens?

Well, I was part of the Ma Ying-jeou administration in 2014. I joined the cabinet…

I should say, the previous administration, which is a person who was with the KMT, as opposed to the DPP, who are currently in power.

Exactly. So, and I’m not alone in this, right? Our trade negotiation minister, John Deng, was minister of economic affairs also in the Ma Ying-jeou administration. So, I think at this time, there’s more independent cabinet members than members of any party in the cabinet. So, when I say credibly neutral, this is what I mean.

Is it difficult to remain neutral?

Well, yes, it is quite difficult when, especially closer to an election, when people naturally drift toward one presidential candidate or the other. There’s a tendency to say that, you know, this is like, not just is my vaccine better, but yours is actually toxic. So, something like that happens close to any election.

On the other hand, by just focusing on defending against foreign cyber-attacks and working with the democracy network in joint defense and resilience, we’re many times removed from partisan politics. So, I guess our mandate makes it easier for us to remain neutral, much like the cyber command or the army or navy or air force, which is supposed to protect the entire country, not just one party.

And do you, because when we think of cyber-attacks and some of the figures that you quoted at the beginning, they are quite shocking in a way. And do you have the funding and the people to be able to combat all those attacks that are coming, which I imagine are going to grow just as digital presence does around the world?

Yeah, that’s a great question. As I mentioned, our ministry started right after Pelosi’s visit, and the missiles flying over our head and the cyberattack that we’ve never seen to the degree of, so yes, we do get sufficient funding in terms of cyber resilience.

In fact, we invest in multiple providers for satellite connections. We not only work with SES, which is this European mid-Earth orbit supplier, but also OneWeb, which is UK, along with many other providers in public clouds like Google and Microsoft and hopefully soon Amazon that provides the resilience.

So, when every time there is a cyber-attack that breaks or disrupts the service of one vendor, of one subsea cable, of one data center, we’ll have two or more that’s in place that can still provide availability. Of course, it’s going to be more costly than relying on a single vendor, but I think all the four parties in the parliament now see this as essential investment.

And does it ever keep you awake at night, the worry of it?

No, actually, last August, when there was this huge cyber-attack, I told the Director General of the Information Management Department, now head of our administration for cyber security, I told Ms. Hsieh that there’s only one KPI I will monitor you, which is how many hours…

Key Performance Index for our listeners.

Yeah, I’ll do this again. So last August, at the height of cyberattacks, Ms. Hsieh, who was in charge of information management and now leads the administration for cyber security, I told her that there’s only one key performance indicator I will monitor, which is how many hours you slept every night.

Because that’s the battlefield. If the psyops, if the cognitive warfare, if everything renders us unable to sleep, then we don’t get to learn from what transpired during the day. Humans learn by four or more hours of sleep in which the short-term memory is internalized into long-term memory. And if we don’t get hours of sleep every day, then we don’t get to learn from those attacks, which means that novel ways of attack will always catch us. But if we sleep for eight full hours, we are able to come up with very creative strategies.

So, you sleep well?

Yes, I slept last night nine hours, so pretty well.

I’m a big advocate of sleep, so I understand your explanation on that. Just one or two more, and then I will let you go.

Sure.

I learned a phrase from you looking at some of the talks you gave, which was fork the government, like fork in the road or a piece of cutlery, a fork. Can you explain how that plays into Taiwanese, it’s not really Taiwanese politics, is it? Can you tell me how that plays into the Taiwanese digital landscape?

Sure. Fork in software development means taking a pristine copy of something that’s working into your local copy and steering it to a different direction. So, not reinventing anything, but rather steering something to where you think would look better.

That phrase came because I was a very early advocate and participant to the g0v or Gov Zero movement, which looks at the government websites in Taiwan, which is something that Gov.tw, and make a carbon copy of that website posted as something that g0v.tw. So, it looks almost the same.

Instead of the O, there’s a zero.

Exactly, right. So, for example, if people like to learn about budgets but find the PDFs that was 10 years ago very difficult to understand and impossible to interpret and there’s no way to press like or unlike on particular budget items, they can change an O to a zero and went to budget.g0v.tw, which provides the visualization of the budget and allows to interact and such.

So, the idea is that whenever people think there’s something that the government is not doing well, it’s not protesting. It is a demo of taking, for example, the government had a PDF listing of where to find medical grade masks in early 2020. But the Gov Zero people took a carbon copy and made a navigating GPS map or whatever based on Google Map and also OpenStreetMap where you can see at a glance where are the remaining mask near you and you can do a lot more, actually. Contact tracing and vaccination and so on were all forged by the civil society groups.

And so, the great thing about software is that what’s forked can be merged. So, when we saw that the Gov Zero way of doing contact tracing is much better because it just involves scanning a QR code and pressing send to SMS number and that’s it. You don’t have to type anything. Then we just adopted as the national recommended way to do SMS-based contact tracing.

And so, by forking in the civil society, they actually serve as research labs of sorts so that the Taiwanese Ministry of Digital Affairs now can simply merge the best of breed from the civil society.

Interesting. Um, you did write your own job description.

Yes. So it’s really a prayer or a poem. So that was in 2016 because Taiwan never had a digital minister. So, people keep asking me, what does a digital minister do? And because in Mandarin, as I mentioned, ‘shu wei’ means plural, not just digital. So, I had this wordplay and wrote a job description based on the idea of digital as plurality.

So, it’s very short, I’ll just read it to you: “When we see internet of things, let’s make it an internet of beings. When we see virtual reality, let’s make it a shared reality. When we see machine learning, let’s make it collaborative learning. When we see user experience, let’s make it about human experience. And whenever we hear that singularity is near, let’s always remember the plurality is here.”

Very nice. One more, just thinking about AI, this is the last one. Where do you see it going? You’ve talked about, I think you see AI or the internet really as a tool that can be used. It’s not good or bad, it’s just a tool. But, because you can probably look into the future… [laughter] I mean, one of the top 100 people in AI. Where do you see it?

Yeah, so with the recent actions like the Bletchley Park Summit and so on, we’re now seeing a convergence between what’s used to be two different camps, the progress camp and the safety camp. They are now merging into what’s called a race to safety. Meaning that the top AI labs, instead of investing 30 times more researchers to capability and power compared to safety and care, they are now seeing that by involving everyone through alignment assemblies, collective intelligence, they can meaningfully engage everyone to work on care and safety and alignment and so on.

So, by enlisting more participation from the people, we don’t have to see progress and safety as a dilemma. We can have a bit of both and a lot of both by people participating in societal risk evaluation together, what’s called open red teaming and things like that. So that through participation, we can converge the values of progress and safety to a race to safety.

A race to safety.

As a follow up, can I just ask, it also depends on the society where this AI being developed… If you think about in PRC, there is a very different approach to AI. Minister, here AI is being developed in a democracy, but if it’s being developed instead in a country, autocratic rule, for example, couldn’t it lead to a very different product?

So, the way I see those adopters going, is that it will become like personal computing. People can run in their phone, in their laptop, even entire language models and image models and other models tuned to their liking. And like any personal computing innovation, this empowers people, even if they live in an autocratic regime.

Unlike internet censorship, which means that you keep connecting to Wikipedia, basically, and the firewall can block your access at any given point, this is more like Wikipedia compressed into an avatar that lives on a flash drive. You can just keep querying on your phone and computer without the need of an internet connection.

And so, I think as long as we make this alignment inexpensive and can run on commodity hardware, it will also empower people who are suffering from the lack of internet freedom in autocracies.