Audrey Tang on Civic AI, Democracy, and Agency

Audrey Tang is Taiwan’s former Minister of Digital Affairs and a longtime civic hacker working at the intersection of technology, democracy, and collective intelligence. In this written interview with Unrubric, we asked her about civic AI, disagreement, open models, democratic participation, and what agency looks like for young people.

1. You left school at fourteen. Did that change the way you think about authority, or about who gets to decide what counts as knowledge? I’ve also thought about leaving university early someday to build something full-time. How did you know you were moving toward something you wanted, rather than simply away from a system you didn’t like?

Audrey Tang: What changed was not my view of authority in general. It was my view of where knowledge is made. Around 1993 I encountered Project Gutenberg, a volunteer library of digitised public-domain books, and it became my primary education. Once I started learning that way, the textbooks lost their attraction, because I could finally watch how knowledge was generated rather than receive the result. When the Web arrived in 1994, the researchers working on those same classics were online at the same hours I was, posting preprints with their email addresses attached. Across the internet nobody knew I was 13. They treated me like one professor writing to another. School would have taken about 10 years to bring me into that conversation, and another 10 for the newest work to settle into a textbook. At 14 I stopped going. On how I knew, I am not sure knowing is the right word. What I can point to is a record. I was already doing the research and already writing to people, and they were already writing back without knowing my age. Away is a feeling, and feelings are honest without being evidence. Toward leaves a trail outside your own head. That is a test of the work, though, and not of anybody's decision about school. What I left school for was not freedom from institutions. It was Project Gutenberg and the preprint servers, places where nobody was obliged to answer me and where being wrong in public was the ordinary way to learn. The circumstances around my case do not transfer. Leaving school then was neither lawful nor ordinary in Taiwan. My head of school told me to stop coming and said she would cover for me, meaning she would falsify the attendance record. My parents' first worry was a gap in my knowledge that would not show for years. Taiwan has since introduced statutory self-learning provisions, which is a better arrangement by a distance. I say about Taiwan that it is not a model to copy-paste; it is just a demo. A biography is the same.

2. Digital democracy is meant to give people a voice, but not everyone has the same access, skills, time, or confidence to take part. Even if a platform is open to everyone, the most active people may still have the most influence. How can we avoid that?

Audrey Tang: The concern is right, and the most useful thing I can offer is our own numbers rather than our own theory. In March 2024 Taiwan ran an alignment assembly, a kind of citizens' jury, on information integrity. The government texted 200,000 randomly selected numbers, received 1,760 valid responses and convened 447 people in 44 groups of about 10. Among those who came, 88.3 percent held a college degree or higher and 92.7 percent used a computer or the internet at work. Only 2.3 percent came from the eastern region, and nobody from the outlying islands of Kinmen and Matsu attended at all. As the process ran online and in Mandarin, it likely undercovered noncitizen residents, including hundreds of thousands of migrant workers. Random invitations cannot establish representativeness on their own. Weighting corrects variables among the people who arrived. It cannot supply the experience of anyone absent. Two design commitments follow, and neither is a platform feature. The first is to invite rather than to wait. An open form measures who had the evening free and the confidence to use it. A random invitation measures who we actually asked, and we pay people for the hours it takes to read the briefing and sit in the room. The second is to stop rewarding volume. On Polis, the platform we use to map where opinion actually sits, there is no reply button and no way to repost, so nothing travels by dunking on a neighbour. Statements rank by agreement across opinion groups rather than by the size of the largest group, and a strong rejection from a coherent group pushes a statement down. In the assembly rooms, which ran on Stanford's deliberation platform, each person listened to nine other people while software kept the queue and the clock. None of this makes a process ungameable, and coordinated voting can still lift a ranking. It does remove the reward for sheer volume, which is the part a platform actually controls.

3. A lot of your work is about helping people find common ground. But sometimes people disagree on very basic values. How do we know which disagreements we can live with, and which ones actually need to be resolved before people can move forward together?

Audrey Tang: I do not sort disagreements by how deep they are. I ask what decision is actually due, and who is owed an answer when it is made. Where no decision is forced, the standard is rough consensus, borrowed from the engineers who built the internet. It does not mean everybody agrees. In the 2024 assembly, most items ended near 85 percent saying this is a good idea, with the remaining 15 percent saying they could live with it. In Taiwan we say 雖不滿意但可接受, not satisfied but acceptable. What a good process looks for is uncommon ground, 罕見共識, the overlap people did not know they had because the differences were louder. In the 2015 ride-hailing deliberation the two camps never stopped disagreeing about whether the platform was extractive or liberating. They agreed on insurance, on registration and on not undercutting metered fares. We wrote that down and left the rest of the argument standing. Where a decision is forced, a narrower settlement can answer it without converting anybody. In 2017 our Constitutional Court ordered marriage equality within two years. A referendum on a straightforward rewrite of the Civil Code then failed in 2018. Many who objected worried about extended families being bound together by marriage, while most young people who planned to marry had no wish to bind their families either. The 2019 act legalised same-sex marriage while exempting kin, so no wider family was tied in. The rights question was answered inside the deadline. The disagreement about what a family is went unresolved and did not need resolving. One more thing, as it is easy to hear deliberation as a process that always ends in agreement. In 2024, after hearing the experts, our participants cut their support for mandatory algorithm disclosure by 27.5 points to 55.6 percent, and the measure did not proceed. It had been a favourite of the conveners. On a neighbouring item the same room moved the other way, raising support for detecting AI-generated content by 2.6 points to 89 percent. They were weighing each question on its own merits rather than following a script. Agreement is a signal, never a verdict on truth, justice or the views of people who did not take part.

4. You’ve talked about Civic AI being smaller, more local, and controlled by communities. Would you ever choose a less capable AI if it gave people more control over it? Could a less powerful AI sometimes be better for society?

Audrey Tang: I would not choose a weaker model — indeed, Thomson-1.0-Small, a model I use daily, is a stronger model in journalistic writing than many cloud models. Capability and control are two different things, and the interesting cases live in the gap between them. You quoted my line about technology on tap and not on top, so let me stay with that image. What decides whether a community governs a system is not the model's size. It is local evaluation capacity, the ability to test a system against your own cases, adapt it and exit when it stops serving you. The frontier labs keep getting better at evaluation. Nearly everyone else is getting better at deployment. That gap, rather than the capability gap, is the one I find worrying. For my own work the local option is not the weaker one. I keep what I call a Kami, short for Knowledge Artefact Management Intelligence, borrowed in shape from the Japanese folk idea of a small spirit attending one place and never the whole world. It is bounded to a single purpose. Its memory and source ledger stay on the laptop, and it runs on a machine with 16 gigabytes of memory. Given access to local material it can be faster and more useful to me than a remote frontier service. The provider is replaceable. In its house rules, retirement is not poetry, it is a contract clause. Then there is the supply question. Fable 5 launched on June 9, 2026. Three days later a U.S. export-control directive led Anthropic, which makes both models, to suspend it briefly, as users' nationalities could not be verified in real time. Capability that can be withdrawn overnight is not capability anyone holds. Access without due process, fallback capacity and accountability is not sovereignty; it is dependency with better branding. So yes, a smaller system sometimes serves a society better, though never because it is smaller. Because it can be inspected, corrected and interrupted. Interruptibility is not a flaw in democratic technology. It is one of its constitutional virtues. The test is whether the people who inherit a system can fix it when it breaks.

5. If AI helps a group find common ground, it still has to decide what to summarize, what to highlight, and what counts as a useful connection between different views. Who should make those choices, and who should be able to challenge them? Can AI really help guide a discussion without also shaping it?

Audrey Tang: It cannot do so by itself. Every act of facilitation shapes a conversation, and a human facilitator shapes it simply by choosing who speaks next. What matters is whether the shaping stays visible, reproducible and contestable by the people being summarised. That happens in three stages. In the first stage the machine decides very little. On Polis the participants write the statements and vote on them, and what we call group-informed consensus is arithmetic over that vote matrix. A statement rises when it draws agreement across opinion groups and falls when a coherent group rejects it. Whether something counts as a bridge is settled by the voters, not by a model holding a view about bridges. The second stage is where language models arrive, drafting themes out of thousands of contributions. That gives the summariser real authority, so it has to come with a way to check the result. My conditions are the ones we would demand of a vote count. Publish the steering prompts, meaning the instructions given to the model. Open the code. Pin the model so that identical input returns identical output. Publish the sources as a tamper-evident package, so that anyone can download the whole apparatus, change one source and see exactly how the report changes. A sensemaking result nobody can replay is an election result nobody can recount. The third stage is the power to say no. Participants and facilitators need the standing to inspect, revise and reject the categories a model proposes. Without it, a model provider ends up standing between people and the meaning of their own words. Whatever does surface then has to reach a named institution that adopts it, explains itself or refuses in public with reasons that can be challenged in the next round. The same question is worth asking of any publication, including yours: when we summarise or translate somebody's words, how do they correct us, and how do readers see the correction? The role of these systems is to surface the quiet voice. The AI does not decide what is fair or who bears the cost. That stays with the people, and those who will carry the consequences should help author the test.

6. Young people now grow up with systems that decide what information they see, what gets recommended to them, and even how ideas are summarized or generated. In that kind of world, what does it actually mean for a young person to have agency, rather than just choosing from options that someone else has already shaped?

Audrey Tang: Let me start at the smallest scale, as I am not exempt. Some years ago I noticed that the phone was scrolling me rather than me scrolling the phone. My corrections were unglamorous: a stylus instead of a thumb, an old keypad phone after working hours, a break every half hour, eight hours of sleep. Attention is the one resource nobody can delegate. Personal practice matters, and so do the rules around it. Three of those do more for a young person's agency than any habit of mine. The first is what school trains. When we rewrote Taiwan's basic education curriculum, we replaced literacy with competency, because literacy is receiving information while competency is closer to training a muscle for collaboration. The 2019 rollout made curiosity, collaboration and civic care the three pillars. By 2022 the ICCS international civic-education study placed Taiwanese 14-year-olds top of the world on civic knowledge, which I read as young people feeling that their own muscle can steer society before they turn 18. The second is the right to leave without losing everybody. Utah's Digital Choice Act came into force for data portability in July 2026, with real-time sharing of the parts a person chooses following in July 2027. It works the way keeping your number does when you change mobile carrier. Choosing among options somebody else shaped is thin. Being able to take your people with you when you go is thick. The third is standing before you qualify for it. Every Taiwanese cabinet minister is advised by someone under 35, a system that has run for more than 10 years, and several of my own reverse mentors were under 18. Our vTaiwan process, Taiwan's online public-consultation system, has no age limit at all. Some of the most constructive agenda-setting we saw came from students who would not be eligible to vote for years. Underneath all three sits one distinction. Send a robot to the gym with your membership card and it earns a perfect score while your own muscles atrophy, and you miss the friendships you would have made there. A good system works as a spotter instead. My working definition of agency is the one I learned at 16: 創業的定義，就是不等別人來錄取你. Starting something means not waiting to be recruited. That has not changed in the Age of AI.

The singularity may be near, but the plurality is already here. Lately I have been adding one line to that: we the people are truly the superintelligence. You are already in that sentence.