Democracy is a technology — but it’s a social technology we can all experiment on, and it can produce new inventions.
Artificial intelligence may be the most powerful technology of our time. But the debate over what it means has hardened into two distinct camps: those who see unprecedented opportunity, and those who fear economic disruption, social disorder, even human extinction.
Audrey Tang, a world-renowned innovator and Taiwan’s former digital minister, believes the way out begins by designing AI platforms that prioritize agreement over outrage. Taiwan used that to rebuild public trust. Now Tang is taking this approach to governments and industry, urging them to see AI not as a race, but as—
Well, Audrey, great to have you coming to us today from Taiwan. To make sure I don’t blow this interview, I brought into play Amy Tong, who, as you know, is the chair of the Ballard Partners group on emerging technology and AI practice. And she has about four decades of experience helping drive the digital revolution in California, which is, you know, Amy, kind of state of the art when you talk about best practices.
I would hope so. I would hope so. But I would say three decades.
Oh, did I get it wrong?
Four decades! Oh my goodness. I would have really good skincare if I’m four decades into this. [laughs]
Well, we just — Audrey, we just had a show with Amy, we taped yesterday, and when she talked about all of her long experience, I was worried that people were doing the math. And I said, well, of course, she started when she was two.
That is true. That I will go with. [laughs]
Yeah, I also have three decades of contributions to civic tech and free software. And I did start when I was 14, so… [laughs]
Well, what are you going to do in the next part of your life? That’s what I want to know. You’ve done so much already, Audrey.
So, there’s an increasing tension between what I think you referred to, Audrey, as the doomers and the bloomers, right? About whether you want to embrace AI and all this new technology as the harbinger of prosperity, or it’s the beginning of the end. How did we get to this point?
Well, I think part of the reason is the metaphors we use. Now, I’m academically based in Oxford, and Oxford, as you probably know, is home to this idea of Superintelligence, which back in 2014 was a very popular book. And the people I’ve met today, when I talk to them about AI, they’re both worried and hopeful at the same time — they do have this mixed feeling. But then the loudest voice, eliciting this idea of superintelligence, really filled the vacuum.
So on one hand, we hear that the AI system will, you know, hack all the sandboxes, all the cyber air gaps — which, by the way, is no longer fiction. It actually happened the day before we recorded: we learned that OpenAI’s unreleased model did hack Hugging Face.
But then on the other hand, we are also promised that it will fix all the problems in the world. And I think what we really need to do now is to focus on the practical questions. Like: how should we steer this together?
Because when the book Superintelligence was written, the state of the art of AI technology was recommendation engines — and they really are maximizers, just trying to keep us glued to the screen, engagement through enragement, and so on. So I think by focusing on the technology as it actually is, not what people imagined twelve years ago, is a very good start.
But Audrey, we’re kind of new into all this, right? So isn’t it kind of natural — since this is representing change across the board — that it’s going to take us a while to figure it out? About how we feel about it, and how we interact with it, how meaningful it is, how perhaps dangerous it could be. We’re still new into all of it, right?
Yeah, it’s completely natural. I mean, every general-purpose technology takes time. The printing press took generations before society worked out how to live with the power of Gutenberg.
And the difference, I think — what many people feel now — is the speed that things are happening to us. And we may not have generations; we may have to compress that learning. But the good news is that the language models are also very good at getting people to see each other’s perspectives. I’ve used models a lot to translate across disciplines and cultures and so on. So you can actually just ask people, in natural language, at scale: “How should we steer this together?” And then just point that to the language model and say, well, train your next generation to be more like that.
You mentioned speed. How do you feel about the fact that many people are looking at this AI competition — you know, whether it’s from the tech industry or policymakers — as a single race going for a single finish line, rather than a mission to achieve, you know, the power of human intelligence?
Yeah, that’s a great frame. I’d say the question is whether this is a race, like a space race, or a mission, like a space mission. A race has a finish line that nobody actually ever reaches, because superintelligence — the finish line — is defined as a technological singularity, literally something beyond human comprehension. So you can’t even reach that finish line. We can’t even reason about it.
Yeah, but a mission has milestones that everyone can contribute to. And if you lead a mission, others want to align with you. And if you say, “I want to win this race,” well, your competitors simply fear you — or maybe sabotage you. So it’s a completely different valence. To lead a mission, I think, is vastly preferable to leading a race.
So we have kind of reached, at this moment at least, a stalemate of sorts, right? You have the loudest voices on both ends that are trying to get attention. You have everyone in the middle trying to figure it out. But it feels like we’re not sure what the next move is going to be, and what the next move should be. You see it the same way?
Well, I think we have seen that many people can reach this uncommon ground in face-to-face settings. In Rhodes House in Oxford this March, we invited people who are worried about extinction risk, people who are accelerating opportunities, people who care about care ethics — and aligned them: educators, developers, everyone in the same room, at Rhodes House in Oxford.
And then we asked people to go online while they’re having these table conversations, to record their statements and also upvote or downvote on each other’s statements. But because there’s no reply button, and there’s no quote-tweet or dunk button in the Polis system that we use, there’s really nowhere for a troll to grow. And people very quickly saw the sense-making results: even though they may self-identify as a doomer or a bloomer, they can actually all agree on some very, very basic things.
Everybody agreed that when people teach each other — the new generation — these kinds of technologies, we should always see that it is steerable by the younger generation, instead of forcing the younger generation through disruption, forcing them to adapt. That’s a very good idea.
And people also said that in government services, in any case, people should always be able to reach a real person when dealing with that service, even when AI handles most tasks. So, always a way to escalate to real humans — and that had like 97% support.
So there’s actually not that much division. It is just that the loudest argument gets amplified on online social media, because of the engagement-through-enragement algorithms.
Can you elaborate a little bit more? Because, you know, on uncommon ground, right — versus, you heard about people needing common ground, or like a completely different approach to doing things. And it seems we’re stuck in this polarized world of: you agree, you are in one camp; you disagree, you are in a completely different camp. You just talked about maybe there’s a different way of treating people’s different opinions.
That’s exactly right. I think the misaligned recommendation engine circa 2014 really figured out how to reward-hack our attention, because people feel that it’s much easier to dunk on someone — the affordance rewards that dopamine hit. And then the platforms started featuring those dunking posts more, because it elicits more interaction, more addiction, and so on. And then it learned to sort the other things to the kind of below the fold, so to speak, so it doesn’t get people’s attention, even though it is actually the common ground. So the common ground became more and more uncommon.
So this is what you can find when you stop asking people to compromise and start asking them what they already agree on without knowing it — like, “Oh, people on the other side also think that.”
And so, for example, on X.com, we now have an algorithm for uncommon ground called Community Notes. If you see a post that is noted by Grok or a contributing editor, then everybody can upvote or downvote it. And again, there’s no way to dunk or reply to a Community Note. So if you don’t like that note, you can downvote it, you can propose something better, but you cannot fight with it.
And then finally, the notes that get shown are the notes that get the unlikely upvote from both sides of the political aisle. And that’s really the only trick. But then it’s much more trusted than any independent third-party fact-checking organization, because it is kind of literally the will of the people, shown adversarially: both sides try to attack these notes, and when both sides are like, “Okay, I can live with it,” then that note tends to be really good.
So, you know, what’s fascinating about this — and we had a prior conversation, Audrey, about your experience in Taiwan. You joined the cabinet there in 2016, but two years before, a public poll — or maybe many public polls — measuring public support for government was down to 9%, right, following the Sunflower Movement. And now you fast-forward and there’s huge support for government in Taiwan. And yet that support was not conditioned on everyone uniting on everything. It was conditioned on the very thing we’re talking about, which is: come to the table with your ideas, your beliefs, and let’s figure out what to do with that — as opposed to who’s going to win today, which side is going to win today. It’s not A versus B; it’s A and B.
That’s exactly right. And I think what people see with disruptive technology such as AI… the peak slop that we’re all living in now affected Taiwan a couple of years ago, like in 2024. We scroll on Facebook, on YouTube, and we always see Jensen Huang, the Taiwanese CEO’s face, put on a deepfake scam campaign. So the fake Jensen would try to give you investment advice or some cryptocurrency. And when you click, Jensen actually talks to you and really sounds like him. But of course it’s not him; it’s just a deepfake.
And so, in Taiwan, because our internet freedom is top of Asia, we really cannot do censorship. That is just not a policy that people would accept. So what to do?
Again, we use the same method. We sent 200,000 text messages to random numbers around Taiwan, asking what should we do together. And thousands of people volunteered. And through random stratified sampling — like a rigorous poll — we chose 447 people, and then they met online in tables of 10, each person looking at nine other people, with only one ground rule: you have to get your idea to be considered livable, right — the rough consensus of your table — in order for your idea to bubble up.
So for example, one table said, “How about let’s declare all advertisements as probably scam online, like cigarette labels, and then somebody digitally signs it, and then we can take it down?” It’s a good idea.
But some ideas that initially sounded good — for example, one table initially said, “Let’s compel all the platforms to open up their algorithm, so everybody can see exactly what they’re profiting from those scam advertisements. Are they earning more per click, and so on?” And that initially had like 70 or so percent support. But after deliberation, everybody understanding that the fraudsters could now then game this algorithm if it’s open, then it fell by 25%. And so we did not take that.
But a few months after that, it did become law. And so throughout 2025 — last year — according to Reuters and the ministry’s numbers, scam fell by more than 94% in terms of investment in deepfake scam on Taiwanese social media, thanks to the law that was crowdsourced this way.
So I think that’s a very clear illustration. That is to say: when you poll people individually, they’re all NIMBY, NIMBY — the extreme positions. But in tables of 10, they’re all MIMBY: maybe in my backyard, if you do this, if you do that. And you get very thoughtful ideas.
Taiwan has roughly a population of 23 million people. So, the United States — it’s far larger, you know, deeply, deeply polarized, and in many cases a lot more decentralized. How do you see whether this is something that can be scaled, and specifically when you come back to AI as an example, right? Right now people are feeling like, hey, things are happening to us — you know, whether it’s policy level, or the speed of the tech companies rolling out AI. Like, is that something that can be helpful here?
Well, I mean, the decentralization is a feature of the US. You have 50 labs of democracy, and some of them, like California, are frontier labs. And so, in California already, we can go to engaged.ca.gov and say, “Okay, I’m a Californian, and I feel that my work is being impacted by AI. I want to talk with people who feel similar to me.” And maybe for different ideological reasons we would support the same package of proposals — and then ideas move forward and then inform real change.
And I think this is now being taken in many, many places. I also advise the Bloom project, which is starting very similar conversations on a community scale in the state of Utah. And again, we have seen bipartisan support there, and also in Central Oregon. And more regions are all coming soon.
And so I think you don’t start by top-down saying all of the US should be in the same room. You start where people are — with their communities, their churches, their sports clubs, their civic classes — and then scale across, not scale up.
Okay, I got to ask you this, as you’re going through that Utah note. As you look at the United States, where we hear all the time — Amy, right? — about red states and blue states: that this works regardless of color?
I think this works wherever we have disagreement. The only fuel we require is people disagreeing with each other. And so, in a sense, polarization is fuel, and the US has an abundance of that fuel.
What this kind of platform cannot do is to convince people to care about these issues. So the more debate there is, actually the better the quality of the conversation.
In Kentucky, in Warren County, we have advised a project called What Could Bowling Green Be? — What Could BG Be? And I think around 10% of the population around that town has participated in exactly the same Polis conversation — upvote, downvote, and so on. And then we used language models to aggregate the more than one million interactions in Bowling Green, around arts and culture, infrastructure, transportation, economic development, and so on.
And you just see people who initially thought they’re very polarized — because this area is maybe more blue, the surrounding area more red — and so they start with, like, very strong sentiments. And then over the course of weeks, because only the bridging statements go viral, people start seeing more and more viral statements that speak to both sides of the aisle. And so the wider the distance, in a sense, the more creative the statements.
You used the word “civically enabled.” And when I heard that, it made me feel good, because I think people want to be reconnected. I think there is a sense, certainly in this country, of being disconnected: disconnected from government, from leaders, disconnected from institutions, disconnected even more fundamentally from each other.
And if this could be, as you call it, a bridging mechanism where we find a way to find the best in each other — it is something we have fundamentally not tried in any concerted fashion in this country for certainly the last five to ten years. And Amy asked the question: can you scale this up, you know, from a Taiwan, which is 23 million people, to the United States of America, which is over 350 million people? Why not? What’s stopping us from doing that, Audrey?
And I’m happy to report, not only California — which is double the size of Taiwan — already now adopting this. I think soon to be a permanent part of the Office of Data Innovation. Yes, I’m just reading the bill moving through the statute. So hopefully this becomes a permanent part of the entire state apparatus.
But we’re also seeing the specific form that I described in Taiwan — we call it Alignment Assembly; it’s more well known as a citizens’ assembly online — it’s now also being taken up. I was just advising CT.MOV, Connecticut Citizens Assembly, on property taxes. And so again, we use this kind of platform to do a maxi-public agenda setting, and then randomly choose people to meet in person, this time for a mini-public, deep conversation. And it really worked very well.
And so I don’t see anything preventing this from happening. I think mostly it’s a lack of common knowledge — like, not many people know that the neighboring state or the neighboring city are already doing citizens’ assemblies.
How did you get into all this, Audrey? I mean, because you’re on, like, the top-10 list of some of the most amazing technologists in Taiwan. Obviously you’ve given TED Talks, you’ve spoken to groups around the world. What got you started, and what keeps you going in a field that is relatively brand new?
Well, I got into this, I guess, when I was four. I was diagnosed by doctors — to me and my parents — that I only had a fifty-fifty chance, because of my heart condition, to survive until surgery. I did get a VSD surgery when I was 12. But for eight formative years of my life, I went to sleep every night feeling like a coin toss: if it doesn’t land well, I just don’t wake up the next day.
So I developed this kind of defense mechanism called “publish before perishing.” So I would document everything I learned that day — first in cassette tapes, floppy disks, larger and then smaller, and finally the internet.
And I found out that on the internet, if I publish my half-formed ideas online, it’s actually making me a lot of friends. It’s a civically enabling experience — because if it’s perfect, people just like it and then they swipe away. But if you post a typo or some misinformed thought on the internet, everybody shows up and says you’re wrong. And that’s how I made my friends this way, because I really don’t have time to be perfect.
And so I think when I dropped out of high school, when I was 14, I wanted to investigate why people so readily trust each other online — called “swift trust.” And in some not-so-well-designed spaces, why do people lose trust, even among people who were friends online — “swift distrust.”
So, applying my research, I participated in the first dot-com wave in ’95, co-founding a kind of large unicorn at the time that did C2C auctions, a little bit like eBay and many other things. And then I went into the free software community, always with the same idea: that we should be able to turn the internet into a space where people can readily trust each other with incomplete thoughts. Instead of getting doxxed, we make new friends.
How do we meet the moment now, because this AI is really, really a central topic? What would you suggest that those of us who are dealing and living this on a daily basis do to address this particular challenge?
Yeah, I think your opening frame of “this is a mission” — that’s a really good step one. And for the milestones, we can, you know, just ask the people. Like, at the end of this episode, we can also just ask one question to the audience, and cluster what they send back, and find the uncommon ground with sense-making, and then report back at the start of the next episode. Then the show isn’t just talking about broad listening; it would be doing broad listening, and then broadcasting, and then broad listening, and then broadcasting.
Because the most useful practice, I think, is just demonstration. When we went to occupy the parliament in Taiwan in 2014 for three weeks nonviolently, we always say: “We’re not protesters who are just against something. We’re demonstrators — we’re for something.”
So if the MPs could not decide how is the good process to talk about a trade deal with Beijing — well, the people, a million people on the street, many people online, will, in tables of 10, virtually or in person, every day, inch together on milestones after milestones of what a good trade deal must do and must not do. And at the end of the three weeks, we did converge on a set of uncommon ground, and the speaker of the parliament at the time simply said, “Okay, the people’s version are better. Go home.”
So the idea is to keep exercising it time after time — not to solve it analytically from the get-go.
Well, first of all, I’m very encouraged by this conversation. And, you know, you see a lot of movements around the world right now. I think France recently weighed in, talking about maybe banning social media altogether for citizens under the age of 15 or something. That’s really not the answer, is it, Audrey?
Initially, in Engaged California with Amy, we thought we should start with that question, by asking teenagers: “What would be the acceptable, pro-social use of social media? We allow those, and then we ban the rest.” And then we talked to their parents, in tables of 10 with random sampling, and the governor and his first partner could pre-commit to whatever the children and the parents could agree on, and that could be the direction.
And we got it all planned out. But on the day of launch, we did not launch, because the week before, the wildfire happened. So Amy had to steer the entire platform from talking about social media and teenagers, to talk about how to recover and mitigate and prevent the future wildfires — which she did heroically and very successfully. But I think the original impulse was exactly right.
I think we can learn a lot. This nation can learn a lot from your experience in Taiwan: how 23 million people, over a decade ago, were not liking anything about—
[laughs]
—public government, and people that were seeking to represent them. And then ten years later, they’re united. But they’re not united in a specious way. They’re united legitimately, because they did something that we don’t do enough of in this country, which is to talk to each other.
And allow the space for every resident like ourselves to practice our civic muscle. And this is where the conversation began: the goal here is not necessarily for people to agree or disagree based on their current understanding, but the ability to actually hear each other, and find that common ground from the uncommon ground.
So, Audrey, we really appreciate your time today. I want you to keep publishing. Don’t perish on us — keep publishing, right? Because you’re speaking a language that people are not familiar with, and that they need to hear and learn. Because the more we respect and love each other, I think the more we have to cheer for and to feel hopeful for moving forward, and be able to then see AI and technology as a real asset, ally and partner in it — as opposed to something that many still fear because of the unknown.
Definitely. Yeah. When I’m preparing the next book, at civic.ai, I was struck by this notion of the labs of democracy — laboratories of democracies, as I just said — because it says that democracy is a technology, but it’s a social technology. We can all experiment on it in the 50 labs, and it can produce new inventions. In fact, the town hall, the jury of your peers — many of these things are inventions in America. So what we did in Taiwan was take those ideas and run them at the scale of the internet, and show that they still work. So what you just said — learning from Taiwan — is just really America taking back its own tradition, and now with better tools.
Couldn’t have said it better. Take care of yourself, Audrey. Really appreciate your time. What you really provoked within me is a sense, again, of hope. I feel good about it. The whole thing about being civically enabled — love that. That should be a whole campaign in America, to kind of get us back on the high road, and feeling that we are uniting once again in common purpose.
Thank you. Live long and prosper.
