AI and Democracy: Ambassador Audrey Tang on Plurality in Practice, Transparency and Collective Intelligence

Hello and welcome to the Accelerating AI Ethics Podcast of the University of Oxford. I'm Dr Caroline Green and in each episode we explore bold ideas, innovative thinking and creative responses to the ethical challenges posed by artificial intelligence. Today's guest is someone who truly redefines what's possible when technology meets democracy. Audrey Tang is Taiwan's cyber ambassador at large and formerly its first digital minister. A self-described civic hacker, Audrey helped transform Taiwan's government into a global leader in digital democracy, using open data, participatory platforms and radical transparency to foster public trust. Now as a fellow of the Accelerator Fellowship Program here at Oxford, she is developing a project called Plurality, advancing ethical AI through collective intelligence, a bold vision and one that offers new ways of thinking about how societies can shape AI, not the other way around. Audrey, welcome, it's a real pleasure to have you with us today. Very happy to be here, my first day as an Accelerator Fellow, really good to be in this new space with podcasts and sharing with you all on how to align AI to society and all the other way around. To get started, tell me about Audrey Tang. Audrey, where are you from, what's your story? So I was born in Taipei, Taiwan, when I was four, the doctors told me and my family that this child has a heart defect and has only a 50% chance to live until surgery. I eventually got surgery when I was 12, but for the first 12 years of my life I go into sleep not knowing whether I'll wake up, feels like flipping a coin. And so this gives me two kind of superpowers. One is that I learned Taoist meditation very early on so that whenever my heart beat above a certain beats per minute, I start debriefing and so on instinctively because if I don't I just faint. And the other thing is that I publish before I perish, so every day I take to record what I learned that day, first in tape recorders, those plastic things, and eventually of course to the internet. So I got into the habits of publishing everything into the public domain, into creative commons because if I don't wake up the next day, well, people are going to be able to use it and remix it. Wow, that's quite a story. So as a child when you were going through that, you actually took it as a driver to share with the world what you're learning every single day. Right, it's an existential opportunity. And how has that shaped you into adulthood? So when I was 14, I already went through three kindergartens, six primary schools, and one year of middle school. And I discovered at a time this thing called the internet with browsers, which was just coming up in '95. And also this preprint server called Archive ARXIV from Cornell is still around, is where most people publish their AI findings and so on. And so I started writing professors and they didn't know I was just 14. I actually speak terrible English back then. I had to look up dictionaries, but I was very interested in why people come to trust each other so readily online. In person, it takes hours, days for people to get acquainted to one another. But online was just the right meme, the right hashtag, the right common ideas, the right domain name. So tell each other very intimate things very quickly. And of course, they also break up very quickly, but that's called swift trust. And so I made it one of my research topics. I did the science fair top in the nation. And then I started learning about this research network. People are looking into this kind of new emerging social phenomena on the internet. And so I told the head of my school, Principal Du Rui Ping, I said, "Look, I can do this 16 hours a day doing research on this new thing called the internet. Or I can go to your school, pretend to study for eight hours and then only do eight hours research. I would like your help." And she read my email printouts and said, "Okay, from tomorrow on, you don't have to go to my school anymore." And I was like, "Okay, but it's compulsory education. My family will get fined." And she's like, "It's okay. I'll just fake the records for you." And I tell this story because it's more than 20 years, the prosecutory period is over. But the point is that I learned that bureaucracy is actually innovative. People are willing to actually bend the rules if you say that actually you're of the same value, but we're just taking a shortcut, a better way, a more effective way to accelerate getting to that common purpose together. So Audrey, listening to your amazing story, it strikes me that you are someone who is really fascinated by people, by humans, how their minds work. Can you tell me a little bit more what drives you working with humans? That's a great question. Indeed, when I quit school, I almost immediately put my learning into use by co-founding one of the fastest growing Taiwanese dot-com enterprises, a startup called Inforion. And we eventually got investment from Intel, started CoolBit, which is equivalent of eBay, like C2C Auction and search engines, many other things. And what I've witnessed is that people are much nicer when they're around other people. If you pull them individually, if we put them into a place where they're isolated and just look at snippets of social media posts, enragedments, driven engagement, and so on, then actually people become very antisocial. On the other hand, if we put people in groups like group of 10 and they understand what each of their saying, each of their typing is within a shared context, then people start moving from imbi or nimbi, like very selfish positions, into mimbi, like maybe in my backyard, but only if you also commit something in your backyard. And so I think the social nature of the internet was always what fascinates me and how to design such spaces, such as in C2C auctions, that elicit the best from the people, the most reputable from people, the most ethical from people, that has been kind of my main thing as a entrepreneur. So do you think that humans are better when they are operating collectively? Yes, and this is what we call pro-social media in Taiwan. So a very quick story, in 2014, we peacefully occupied our parliament for three weeks in Taiwan because the president at the time was enjoying only 9% approval rate, so in the country of 24 million people, anything President Ma says, 20 million people automatically skeptical, which was the same situation in many of the Occupy, Arab Spring, and so on movements. But we took a critical different approach instead of calling ourselves protesters, which are against something. We call ourselves demonstrators, which is showing something new. And that something new is anybody who are worried about the trade agreement with Beijing at the time that would have invited Huawei, ZTE into our telecommunication or 4G infrastructure or publishing industry, media, and so on. If you are worried, instead of protesting, saying we shouldn't do this, you can go to one of those corners in the occupied parliament on a facilitated conversation, where people ask each other, "How do I truly feel about this?" And starting from the position of feelings, people's coherence set a very coherent idea by the end of the three weeks. And the speaker of the parliament, Wang Jinping, simply said, "Okay, the people's idea are better than our ideas, so you win, we'll ratify this, go home." And so we became the very rare Occupy because of this small skill, facilitated, group-based conversations actually converged instead of diverged. Wow, I love that. So here you're also bringing in a whole new way of looking at democracy and how technology can help us today to forge new paths in global democracy, and bringing people together. What I also really enjoy about what you're saying is you believe in the good in people. That's the whole way you're coming to your work. Is that right? Yeah, I believe people are good when they're around other people. And so AGI to me is augmented group intelligence. We should develop AI system that augments our civic muscles. Instead of just, "I talk to my chatbot, you talk to your chatbot, we send chatbots to deliberate, make decisions, persuade us," that's very impressive, but it's as impressive as me sending my robot to the gym to lift the weights and you send your robot to run the treadmill. Very, very impressive, but our muscles don't grow this way. So to me, this civic muscle, this relationship building between people who have different ideas, different ideologies even, but manage to find a common ground, I think that is core to democracy. So that's really interesting. I'd love to hear more about that because quite often when I think of social media, when I think of artificial intelligence, it's more anti-human really. It's really anti-positive relationships. We see a lot of bad behavior on the internet. We see echo chambers, so your idea of actually bringing technology to the world to improve relationships between people, to find innovative solutions, that's something really exciting. Can you tell me a little bit more about that? Are there specific projects you've been working on where you've seen the power of that collaboration? Certainly, so 10 years ago in 2015, many of the anti-social corner of social media become much worse than before because they adopted this for you algorithm. Prior to that, people just follow each other and you get a chronological feed of what other people posted that are on your followers network. So people do have common experiences, common knowledge, but after the pivot to the for you feed, they just figure out with the parasitic AI what keeps people addicted to the touch screens. And it turns out engagement is easier through enrichment. It turns out that people when they're isolated in these small screens prefer to see things that are much more sensational and polarizing than that is healing or bridging. And so it's like rewarding your children every time they are mean to somebody else. So, in a sense, this parasitic AI was the first misaligned AI system that turned a neutral platform into a non-ethical platform. And in Taiwan, at the same time, we're experimenting with the other direction. We worked with an open source system called POLIS. But in 2015, we were trying the other direction, the pro-social media. We worked with POLIS, an open source system. When Uber came to Taiwan that year, many people are very afraid they would take the job of taxis that this algorithmic dispatch is going to ignore the professional driver license system. And people would get maybe higher quality service, maybe lower quality service, really nobody knows. And we use the POLIS system to basically ask the entire society, what do you feel about the situation? Because people are experts in their feelings. If we ask what do you think about sharing economy policy, very few people would be able to chime in. But because we just ask, how do you feel? And we show people, once they share how they feel, they can upvote, they can downvote on each other's feelings, but there's no reply button, so there's no room for trolls to grow. And we show people in their avatar, which cluster they belong to. Where are the people that share similar feelings? And that's good for two reasons. First, people understand that actually we do manage to agree with each other on some of those common feelings. For example, everybody felt that insurance is important, everybody felt that while search pricing is fine, but undercutting existing meters is not. And so instead of the polarizing debate, sharing economy, gig economy or whatever, it shows the connective tissue of those groups. And the second is that there's a scoreboard. So the longer distance your ideas resonate across groups, the more bonus you're given. So the most viral ideas are the one that are the most eclectic, that brings the uncommon ground, the rarely discovered common ground among people who initially diverge. So after three weeks, again, we converged online. So we didn't have to occupy any government buildings, but we replicated the same process we did in the sunflower movement and the top nine ideas that crossed the threshold of 85% agreement in all the different groups, regardless of whether they're majority or minority. It unified the society together and we make the law about Uber using these rough consensus from the people. So over the next six years, we would hold more than 100 of those collaborative meetings. And approval rates went from 9% to more than 70% by 2020. Wow, that's amazing. So I'm wondering about, when you tell these stories, how you bring people together collectively in the internet, it's very powerful in terms of beating social isolation, loneliness. But I'm also wondering about groups of people here who are digitally not as connected. So I work a lot with communities, individuals who don't use the internet, and they're possibly, they just don't have the connectivity or so on. What can we do to also reach these people? Yeah, in Taiwan, broadband is a human right. We have the universal service fund so that the telecoms, they can go to even the top of Yushan almost 4,000 meters high and set up broadband connections. If they cannot recover from subscription fees because there's just fewer people there, the other telecom who didn't make such investments must reimburse them on the cost lost. And so the universal service fund ensure that anywhere in Taiwan, no matter how remote the island, how high the mountain, you have connectivity in the form of, I think, just 15 pounds per month, you get unlimited data connection. And so we have been doing that for almost a decade now. And so because of that, no one is left behind. And for people who don't prefer to use the internet, we use the strategy of helping to help us. So the young people in Taiwan, instead of just teaching them digital literacy, which would be about receiving information, we teach them digital competency. So the competency is about finding these ideas together. And for example, when people have disagreement about air pollution, the young people set up air measurement stations and then getting their parents, grandparents to look at the numbers shared together. Or they fact check the three presidential candidates as they were having a debate and that they found some flaw in it, maybe their name appear on national television. And the younger people younger than 18 were actually the most active on our national participation platform. So they would start a petition, for example, saying let's go to the school one hour later, because study shows one more hour of sleep is better than one more hour of study when it comes to grades. And then they convinced the elderly citizens to help on their course, to help them to reach 5,000 signatures, and eventually got the time tempo changed. Of course, they don't just do things concerning themselves. They also, for example, petition for banning of plastic straw on bubble tea takeouts. And those petitioner become cabinet level advices, reverse mentors to minister, become very famous. And then they inspire other young people to start even more ambitious things, like starting a menstruation museum in Taipei and in just two years, remove the taboo about this all together in all municipalities and so on. So with the young people as reverse mentors to senior people, we ensure that even if you are in a rural place and you don't want to talk across the internet with other people, you can talk to your young people who are like ambassadors to the digital world. Well, that's great. So the idea of a human right to the broadband and then that beautiful example of how technology can span the generations for intergenerational activity and bringing people together of all ages, that's very powerful thinking about our changing demographics. So yeah, thank you for sharing that story. I'm going to pivot slightly to different questions. So Audrey, what is AI ethics to you? To me, AI ethics represents the technologies and methodologies to imbue the human society's preferences into the whole cycle of AI development so that we ensure that the AI systems that we make conform to the societal expectations, understand the social context and norms, and is steerable by the society. World AI ethics is like a car that has a very fixed place to go and the only thing you can do is to hit the brakes or hit the gas pedal, that is to say to decelerate or to accelerate. On the other hand, most of the societies do not want a few people in Silicon Valley or in some other big tech to dictate how we should relate to one another. So just like social media, many people would prefer the social media instead of selling our attention to the highest bidder, they would prefer that the social media have more bridging content, more shared experience, and so on. However, without ethics input, the social media network systems are designed in such a way that it pulls people, strip minds the social fabric, and then sell our attention to the highest bidder, and so had it been designed with ethics in mind, they will come up with very different business models. Maybe instead of selling individualized advertisement, they will sell common experiences, curated experience, or subscription-based business models, and so on. So to me, it's not just technical, but also about how the private sector incentive works. So from civic hacker to digital minister to global advocate for democratic technology, it's really extraordinary. What drives you to build systems of radical transparency and public participation? We've already spoken about that now a bit, but just to put that question to you again, what's your main driver right now? Sure. My main driver was the same as when I got into the cabinet position in 2016. As I mentioned, the main problem we try to address is the deficit of trust. People were losing trust in all the vertical institutions, whether it's political parties, or ministers, or journalists, or academic experts, and so on. People much rather would like to trust at a time, people who sound like them, who look like them, who gets more comments, and so on. On the other hand, the social media, as we mentioned, was pulling those influences into extreme positions. So we have two problems. One is that the government, the elites, just don't trust the people enough in order to win back trust, but to give no trust is to get no trust. So the first thing we address with radical transparency is to radically trust the people. If the people see a public service that is not designed well, instead of protesting on the demand side, they can switch to the supply side by demonstrating how to do it better if they feel that contact tracing is not respecting privacy. Because of radical transparency, they can design better contact tracing systems that preserve the privacy, and indeed help Taiwan to last until Omicron, and we never lock down any cities, and we reported one of the best economic growth during the three years. It's not because the elites have good ideas, that's because the people can co-create together. And the other one is to depolarize, to bring people back from the extremes where people hate each other, especially intergenerationally. That was a big issue in Taiwan. So the young people, instead of just lamenting that they cannot outvote senior citizens because of declining population, invite them on the table and then build intergenerational links. And so today, Taiwan is doing very well according to BTI. More than 90% of Taiwanese said that democracy's elites fairly good. On the other hand, no country is an island, not even Taiwan. So even if Taiwan depolarized the society, even if Taiwan rebuilt the democratic resilience, if all our allies succumb to the polarization, to the hate, to the intergeneration of distrust, then this authoritarian notion that democracy only lead to chaos, democracy never deliver, is still kind of like self-fulfilling prophecy. So my main job now is to show that actually democracy can deliver. And just with some tweaks in the social media regulations, in the ethics when we're designing AI, to steer these technologies toward a pro-social direction instead of this singularity vision where it just keeps getting better and better, automate its next generation without the need of humans in training new AI models, superintelligence take off, leaving everyone behind. I think the world needs a better vision than this superintelligence singularity take off, and which is what I call plurality. Yeah. So what strikes me about that is that the Taiwanese government also has the readiness and the openness to listen to people, to embrace that modern age democracy. Is this something you feel we see or you can see also in other governments around the world? Definitely. So I would say first that we learned this technique from many smaller polities. Our ePetition system was from Iceland, Beter de Kavík. Our participatory budget system are from Porto Alegre, Brazil, as well as from Madrid and Barcelona, Kansu and Desitim, respectively. The Polish system was from Seattle. The Lumio system, which we used during the Occupy, was from New Zealand. The thing though is that these technologies previously only worked in smaller polities, 10 million or fewer people. And one of the reasons was that broadcasting was so much cheaper than broadcasting. Once you listen across a wider social distance, you need mediators, you need translators, you need people who inform people of very different backgrounds of what is commonly at stake. And this gets progressively harder the larger your polities are. And now, this year, we already see polities larger than Taiwan, trying out these methods of broad listening and very successfully. I was just in Tokyo, in Japan. So last year, Takahiro Anno, a 33 years old engineer of machine learning, read the book "Plurality" that I co-wrote and decided to run for governor one month before the voting day. Nobody knew him. He has no parties. He simply said, "Let's crowdsource my platform." So anybody can hashtag Tokyo AI and chime in with the platform. They can dial in if they're senior citizens to talk to a voice clone of Anno-san. And you can also dial into their YouTube channel in which his avatar broadcasts 24/7, each and every updates. The uncommon ground that was contributed by the people eventually won actually the first place according to independent ranking on the platform's usefulness, so even better than Koike-san. But Koike-san, of course, won the third time, but she was so impressed that she tapped Anno-san to join the Tokyo government as an advisor to GovTech to help her to do broad listening. And this year, all the major parties, the ruling party, the two largest opposition parties, are all using broad listening to ensure instead of polling people one by one, we can poll people in groups with deliberation where people get to react to each other's ideas and gets much better preferences that are much more about care, about mutual care, instead of just about individual utilities. So in Japan, we're seeing a lot of embracing and extension of these methodologies. Now in California, they also just institutionalized this platform called Engaged California, which again surfaces the uncommon ground. The pilot was about wildfire prevention and recovery in Palisade and Eaton, but now as part of their budget bill, they're now institutionalizing it so that it can talk about many other topics, maybe social media, maybe AI-related governance and things like that, again, with the people, not just for the people. Okay, so that sounds to me like we need plurality ambassadors within our governments. Yeah, definitely. And here in the UK, we also saw the Waves project by Denos that was launched with many local governments where they are using very similar bridge-making technologies called RIMESH to figure out the common priorities of everyday people. And the UK has very strong civic muscles on the community level. And so, yeah, it always starts small, but hopefully it can grow to even national level very quickly because now we have language model that can aggregate those feelings, qualitative findings without hallucination for the first time this year. That's very exciting to hear that the UK government and local authorities are embracing this. And now you're in Oxford with us. Tell us about your Accelerator Fellowship project. What will you be doing? Right. The first deliverable is a podcast which we're recording. And so, we will share many podcasts related to plurality and ethics in AI as creative commons. So, just like my biopic, there's a short documentary called Good Enough Ancestors. It's just 21 minutes, one like four awards now. But the entire footage is open in the public domain. You can go to AudreyT.box to download almost one terabyte of footage. And already, like the Collective Intelligence project is reusing many of those footage to make ethics in AI short films. There's a young adult's novel, there's a manga, an illustrated version in Japan. It's very exciting how we can engage professional communicators and amateur people who are interested in ethics in AI and providing them with raw materials that they can remix and reuse and just make sure that people understand there is actually hope to pivot the anti-social media toward pro-social media and many more besides. So that's the first thing. And the second thing is that I will document the ideas, for example, in Utah, they just passed a law about providing off-ramps between social networks. So next year in Utah, if you switch from, say, TikTok to, say, Blue Sky, instead of just downloading your data, it says that TikTok must keep forwarding your new likes, your new followers, the new contents both ways. So it's like number portability. If you change a telecom, you don't have to change your number because if you have to change your number, new telecom would not be able to compete with old telecoms. When I was a child, the ATM only allow you to withdraw cash if you have a car from the same bank. Again, the new banks have no place to set up ATM until the government stepped in and say to foster competition, we actually need interbank protocol so that you can withdraw cash across the different banks. So it's very good to see Utah and other places now looking at this off-ramps and on-ramps between social networks because then the social network would not be able to trap anyone. You can't switch to more ethical platforms without losing all your social connections, your family connections, your albums, and things like that. And so I will be documenting these concrete policies that are already passed or that's in deliberation and basically share a plurality playbook so that people, especially in governments, that want to steer AI toward plurality know exactly what to do, what kind of laws they have to pass. And finally, I will also share playbooks about how grassroots communities, academic people, as well as practitioners, can just apply those technologies even if they're not governors, and if they're not mayors. They can also use these pro-social bridging platforms. For example, for the AI systems to consult with the people, not just for the designers themselves, but rather the people who are suffering from the overreliance, suffering from bias and so on, and turn those expectations from the people into model specifications. And using those specifications to steer AI, this is a technique called deliberative alignment. And I think this technique holds great promise for people to steer AI toward their social norms. So in September 2025, the Institute for Ethics in AI will be moving into the new Schwarzman Center for the Humanities, which is an amazing new building, all the humanities across the University of Oxford will be meeting in that building. And it's all about bridging, also building bridges with the local community and with the public. So it's going to be an open building, people can come in, they can engage with the academics, the researchers, the staff working within that building. And we've got a lot of very exciting new spaces like a concert hall and places for people to meet. Do you have some ideas of what kind of events, formats we could have in that new building to bring people in and for them to hear about the work you're doing? Definitely. I worked for a very long time as a minister in the Taipei Social Innovation Lab, which is literally a park. There's no walls. You can literally just walk in and see, for example, some self-driving tricycles driving very slowly and interacting with people. These were from MIT Media Lab. And every Wednesday, I open in office hours so people, anyone, can have a conversation with me on the record in the creative commons about the projects they're doing. So the hope when I was setting up that lab is that innovators should not be just in garages without talking to the people. The entire society can participate in the process of creation and come together. And we used also those spaces to hold collaborative meetings in Tainan as well as in Taipei. We held alignment assemblies asking people, "How are AI affecting you? How do you feel about AI?" And with facilitated conversations, we actually tuned our sovereign model, the Taiwan Tide Trustworthy Air Dialogue Engine, with the hopes and fears of people's ideas in Taipei and Tainan. Turns out they have very different expectations about the AI's role in the community. And we used so-called constitutional AI to tune the tide toward people's wishes. And so I'm sure once people understand that it is possible to just come together and maybe walk through a few scenarios of how AI is having an impact on the community and just share how do you feel? How would you like to be better? And almost magically, by the end of the day, you can have a plenary overview of what the different groups of people have felt commonly about. And you can derive policies, you can derive model specifications, you can derive evaluation benchmarks, and so on purely from this kind of people talking and listening to one another. So I would love to hold alignment assemblies in the new building. That sounds amazing. Because I do feel, and this is the work you're doing, right, is to build these bridges between people, amazing, very clever people who are building these AI systems, but often in the silo. And then these systems are available to the broad masses, that you actually need to bridge that gap of knowledge and awareness of what AI is, what the limitations are, how we can use it, how it can be abused. And so these are new methods to really build these bridges between different groups. And so that's very beautiful. We're really excited to have you work with us on that. Yeah, definitely. And it can also lead to new frames of conversation about digital rights. For example, last March, many people in Taiwan noticed a uptick in the deep fake advertisements on social media. People would see Jensen Huang, you know, the NVIDIA CEO at Taiwanese, his image saying that, "Oh, I want to give back to our country. I want to give you some free crypto or things like that." Of course, it's not Jensen, it's deep fake. But if you click, Jensen actually talks to you, very convincing, thanks to NVIDIA GPU that can synthesize deep fake voices in real time. On the other hand, Facebook was profiting from those advertisements because the scammer, turns out, pays more than ordinary small and medium enterprises when it comes to placements. And so instead of the government stepping in and saying, "Oh, the censor, the advertisement," and so on, because the Taiwanese people wouldn't have that. We're the most free in all of Asia in terms of Internet freedom. We simply ask people, "How do you feel?" So we sent text messages to 200,000 random numbers around Taiwan, just asking, "How do you feel?" And they gave us their feedback, their feelings. And we also asked, "Would you like to volunteer on an alignment assembly about online fraud advertisements?" And thousands of people signed up. And we chose 450 people, statistically representative of the Taiwanese population. And in rooms of 10, they deliberate. So the 45 rooms for a long half-day talk about various different measures. For example, one room says, "If Facebook posts an advertisement featuring Jensen, we should assume it's scam unless Jensen digitally signs on it. We should flip the default." The other side, another room says, "If Facebook do this, unsigned advertisements, of course we should find them, but we shouldn't stop there. If somebody gets scammed for, say, $7 million, Facebook should be liable for that $7 million. That's the only way they will comply." Another room says, "TikTok, bite dance, they at the time did not have a Taiwanese office. And so they can simply ignore us when we make them liable. What to do?" And they say, "We should slowly slow the connections so that the service featuring their videos becomes slower and slower to load. And so all their business will go elsewhere." And so all these ideas are on the actor behavior level. They're not content level because they're not censorship. So they're considered proportionate by more than 85% of people, regardless of their age bracket, where they live, their gender, their occupation, and so on. So that was last March. And then we check with the Big Tech in April. We sent in the draft law in May. By July, it's all passed. And so this year, if you scroll in Taiwan, you don't see any fake advertisements anymore. And so this shows that in addition of informing the Big Tech developer, as I just said, this can also very quickly inform the parliamentarians because nobody want to be seen as a pro-fraud party. And so no matter how much lobbying is done by Big Tech and so on, once we show actually everybody agree with these measures that the people came up with, then the Alignment Assembly can also have a policy, teeth, not just suggestions. You co-authored a book called Plurality, The Future of Collaborative Technology and Democracy. Tell me about the book. What are some of the main messages that you want to send to the readers? Sure. The name came from my job description. In 2016, when I first became digital minister, Taiwan did not have such a position before. So the HR asked me to write a job description. Turns out in Taiwan, shu wei means both digital and plural. And so I wrote a prayer as my job description. Very short. It goes like this. When we see the 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 a singularity is near, let's always remember the plurality is here. So that was my job. And so it contrasts singularity, which is this idea of an AI system getting so powerful that it can train its next version with minimal human input. And then the next version can train an even more powerful version with no human input. So it's called a super intelligence takeoff. And by that time, it will leave everyone behind. And human history civilization norm society will no longer be relevant to this new super intelligence. But that vision by default, leaving everyone behind, I don't think is where people want to go. People want actually to remind ourselves that the plurality, the horizontal path is a better path, which is fostering our social differences, our diversity, but using AI systems to bridge these diversities. So we can figure out how to live together. And to me, AGI then means augmented group intelligence, so that any innovations, any invention helps like personal computing, each and every person to feel empowered, every community feel empowered, instead of like mainframes, where you have to submit punch cards for very large data centers to compute, everyone can just fork, remix each other's spreadsheets, desktop publishing, connect the computer together into the internet and so on, and just enjoy a much more horizontal path. So that is the main idea explored in the book. So that's really beautiful, that idea. Because I think when we think of AI specifically, agentic AI, genitive AI, everything, you know, those buzzwords and systems that are out there, it's often about the threat to what we as humans value, whether it's social connection, whether it's our work and the purpose that it gives us in our lives. So you are offering a very alternative perspective. You are offering a perspective that's positive. We are saying, actually, what makes us human is something that technology, that AI can explore, help us understand, expand, is that right? Exactly. So as I mentioned, we use AI systems in our alignment assemblies. So in rooms of 10, instead of a facilitator, the room itself is a facilitator. It encourages quiet people to speak up, it limits disruptions to five seconds or less. It offers real-time transcription so people can see the shared notes, what's going on. And it also uncovers the uncommon ground from people from different ideas to stitch them together. And no human facilitator can facilitate 450 people at once. And even if we have 45 small group facilitators, they cannot come together and mind meld and immediately produce a summarization. However, language models can do that and starting this year was very little hallucination. However, the models that we employ to do so can be open source. They can be very small models. Just to summarize, you do not have to memorize the style of Studio Ghibli or something. And because they're much smaller, they can be run at the edge on phones, on laptops, and they're very explainable in the sense that you can run MRI-like algorithms to detect hallucination and so on. And they're also much more energy efficient. And so these smaller models tailor-made to each and every social situation enhance our capabilities of care, of listening to one another without succumbing to this force, superintelligence notion that it can do everything, know everything, but not very well. So I have so many follow-up questions. But the first one is, how worried are you about human-AI relationships? So, you know, we hear these stories of people falling in love with their chatbot. And we hear of how people are going to have AI friends. How worried are you about that? Yeah, it's just like social media. When designed in a way that it encourage prosociality, it can encourage people who are shy, who are introverts, or who are extroverts, but actually not very good at reading each other's emotions and so on to pay more attention, more care to each other, in which case it's very good. It's like a connective tissue. On the other hand, if they offer a ersatz kind of relationship that actually isolates people from each other, then of course it's very bad, because then we just fall into this addictive part of ourselves, just keep doom-scrolling and so on. And in this sense, generative AI is no different. It can be made to be cooperative in the sense of it facilitates human cooperation, but it can also be made to make people addicted to them so it becomes more and more antisocial. In fact, chatGPT4O was for three days antisocial. It got so sycophant. It agreed with every ideas of yours. Even people who hallucinate suffers from different conspiracy theories. For three days, chatGPT would agree with each and every ideas of you that may be saying, oh, they're conspiring about me. They're reading my brainwaves, the 5G chips and so on. It would just say, oh, you're so insightful. You're the only one with this idea. Don't worry what other people are saying. You're the only one that understand the truth and so on. Of course, Sam Altman, very quickly apologized, says that, oh, that's because we use very quantitative A/B testing. We just say what people feel good, that they engage more with this kind of answers, but we ignore the qualitative reporting by people who are more versed in ethics. They actually raised the red flag, but they were ignored. Then those new systems were rolled out until there's a huge backlash from the reviews and on the social media and so on. I think to me, ethics should not be after the fact. Ethics should be preventative. It should be designed in, because if it's designed in, then it's not just about ad hoc evaluations. It should be part of the pipeline, part of the process. I'm not saying that chatGPT will be as bad as the 10 years of bad dream that is social media, but we do feel that there needs to be a systemic infusion of ethics and deliberation. Indeed, OpenAI said as much that they will engage the society in a much more democratic fashion in order to prevent this kind of syncopacy from happening again. So AI ethics, not as an afterthought, but as the starting point. Exactly. Yes, by design. By design. Help me understand a bit more that concept of care. What does it mean to you? How do you define it? So to me, when we're using or designing any social system, we would like people to know each other better in an empathy, not just sympathy kind of way. Of course, sharing each other's feelings is good. On the other hand, if you cannot mentalize to understand each other's feelings in context, then you tend to just reverberate and trap people into shared misery or echo chambers or things like that, or even outrage and so on. So part of care is the ability to contextualize suffering, to contextualize harm, to contextualize people's interactions so that people can reason also together on how, for example, not to repeat the harm that was done before instead of just seeking vengeance and things like that. So to me, it is an interpersonal skill, but it's also a mindset that says instead of just fulfilling the immediate instincts of us, we mentalize our social settings and settle on a better course forward. And this is in contrast to simply saying you should never do this, you should never do that kind of the ontological kind of command setting. And it's also different from just calculating utilities saying that, oh, I'm a little bit more happy and you're a little bit less happy, but I'm happier than the kind of unhappy you suffer. So in the total, we're actually positive and so on. So it offers a very different calculus instead of just adding things together or just saying that you should do this or shouldn't do this in all situations. It considers each specific situation and figure out how to live better afterward. So what you say, that's a new type of care, like digital care that people will need to develop. They can care for each other digitally, but then also socially, you know, when they meet each other as people. And so I love, I live in London and I love walking down the South Bank and I just love seeing people engaging with each other, playing chess, having fun. How important is it still to really invest also in these social spaces where people come together as not online, but to spend time with each other, to have these types of experiences, you know, to also grow that kind of care that you just spoke about. This is very important. And as I mentioned in Taiwan, digital and plural are the same word, shui. And so digital to me is not about replacing in-person human connections. It is always about finding better ways to link to people who you maybe initially didn't know, like strangers, even though they're your neighbors, or that people who initially you felt are kind of indifferent, not curious, and then discover things actually you do commonly care about, which is again difficult to do in an in-person setting, hard to break the eyes. But SwiftTrust Online means that we can more and much more readily discover the topics that we each other care about. But then from these topics, then we go into in-person strong connections. And one thing about this care is that although, of course, you can be responsible for someone, care means that you take responsibility yourself. You don't just delegate that away to a robot. If you delegate everything away to a robot, you may be still responsible, parents are responsible caretaker, but you actually don't foster your own capacity of care. So there's something deeply personally relational about care that I think the digital is here to reinforce, not to take away. That's very, very nice. It fills me with hope, because often I feel like, are we being stripped of caring for each other? Just because we've got this sometimes digital technology, it seems more like a barrier to meaningful human relationships, rather than it being a facilitator to foster them. So thank you for giving me hope. Yeah, definitely. Turns out it's not just attention that you need, but also awareness as well. There's one follow-up question that I'd like to ask before we close, and that's around singularity going back to that. So you spoke about these very powerful AI systems that don't need human input anymore to develop, and there is a concern about these type of systems. And at the moment, there's no global governance to stop that happening. How do you feel about that? Do you think we need more governance to ensure that these very, very powerful systems don't happen? Well, there are some agreements, for example, nuclear powers by and large have agreed that they do not link those artificial intelligence systems into decision-making system to launch a nuke, which is, I guess, a start. Of course, it's not very comprehensive. It doesn't stop proliferation of any kind. But at least people do see that this is a danger. This is a risk. So nowadays, people are worried, as we mentioned, about overreliance, about addiction, especially for young people. And we're seeing around the world many advocates for age signals, so that instead of the government surveilling everybody using social media, there's a way to keep the privacy, but for each person to signal whether they're over 16 or over 17 years old in such a way that we can design systems around age-appropriate responses. Again, there's some governance mechanisms for that. On the other hand, I think the civil society and researchers can do much more than just advocating for such policies from states and governments. What we can do is, again, taking the horizontal path to show that for particular tasks, for particular uses, actually a generally intelligent system is less energy efficient, is actually less predictable, and is much more hallucinatory than the smaller system that may be distilled from larger systems, but are made for purpose and also much more easily monitored. And so to show a viable horizontal path, for example, in a Paris AI summit, I helped launch along with the open source camp, Yalakwint, and the security people like Eric Schmitz, this idea of a robust and open online safety tool, Roost. And the Roost idea is that everyone can band together instead of waiting for a very large big tech company like Microsoft to detect online child sexual explicit materials and so on in a way that simply doesn't scale, now with proliferation of open source deep fake models. We should actually band together like the cybersecurity community and detect and share our threat indicators widely, and we can legally turn those pictures into text, like grooming text, which is legal to hold onto, and use federated learning and other methods to ensure that we preserve the privacy of people involved. And then in real time, train models that detect in a decentralized fashion and open fashion how to stop such online harms. And so this shifts from just appointing one safeguard organization, pray that it doesn't go bad or get corrupted to a much more resilience-based defense posture. Instead of just playing defense, each and every one of us can contribute to the monitoring, to the solution, as well as to the threat indication. And this resilience mindset involves everyone and leave no one behind and decentralize power. So this is called differential acceleration. We accelerate the non-dual use, defense uses of AI in such a way that's democratic, decentralized, and also defensive. So it's not just the job of policymakers of law and regulation, it's our collective job. Yes, and once the security community and open source community do agree on these measures, policymakers should become much simpler. When I was minister, if the top experts were arguing with each other, my instinct would be, okay, let's wait for another six months, right? But if they do agree, okay, these are the joint investment we should make right now, then for policymakers, it's a very easy check to write. So Audrey, thank you so much for this conversation. From Taiwan's open government movement to your work with the Accelerator Fellowship Program, it's clear that ethical AI isn't just about rules, it's about relationships, participation, imagination, and what you have shown is not just problems, it's actually finding solutions. And that's why I'm so excited to work with you. Thank you. Let's free the future together. Listeners can learn more about Audrey's work and the other fellow's projects by visiting the fellowship website at afp.oxford-aiethics.ox.ac.uk. This has been the Accelerating AI Ethics, a podcast from Oxford's Institute for Ethics in AI. If you enjoyed this episode, please subscribe and share. Until next time, thanks for listening.