A talk by Audrey Tang, Cyber Ambassador and former Minister of Digital Affairs, Taiwan - By Sarah Cao
On 28 May 2026, the Oxford Taiwan Studies Programme welcomed Ambassador Audrey Tang to the Nissan Lecture Theatre. Tang, Taiwan’s first Minister of Digital Affairs and now its Cyber Ambassador-at-Large, took up a question that is both practical and civic. How should democracies live with AI? The right test of any civic AI system, they argued, is whether it increases a community’s capacity to care. The standard is not whether a system is perfect, but whether the people who inherit it can fix it when it breaks, which is ‘how the light gets in’.
From the Four Freedoms to the 6-Pack of Care
Tang began with a familiar text in an unfamiliar register, the Four Freedoms first set out in the 1980s as the terms of a software licence. Read again, they suggested, those freedoms point toward a fuller account of what democratic societies need from their technologies. The framework Tang has been developing rereads those four freedoms as civic muscles, then adds two more — a ‘6-Pack of Care’ that shifts alignment away from top-down hierarchical control and toward decentralised, community-driven arrangements.
Attentiveness comes from the freedom to run a program for any purpose: you can pick up a tool for this particular classroom, clinic or temple, without the designer’s permission, and attend to the needs around you. Competence is the freedom to study and change a program — to know what the system is actually doing in your hands, so you can read it and fix it. Solidarity is the freedom to share copies — to hand the tool to a neighbour, or carry it on a USB stick to a country where the cloud is censored. Responsiveness is the freedom to share modified copies — to fork, so that your fix becomes the next maintainer’s starting point and each generation inherits less debt than the last.
To these four, Tang adds two further muscles for the age of AI: responsibility, which is not a single CEO or president but a reachable, accountable community, on the hook to convene through a defined process on a predefined timeline; and symbiosis, a system that steps back when a community’s capacity grows or its needs change, rather than manufacturing reasons to extend its own usefulness.
Tools, Tang explained — drawing on David Krakauer’s distinction between complementary and competitive tools — either build these muscles or waste them. Send robots to the gym with our membership cards, and the weights go up while our own muscles atrophy and we make no friends along the way. Recommender systems, in this reading, have been competing with the relational fabric of users’ lives for years, and the resulting civic atrophy is visible around us. Close the path to repair, Tang warned, and the capacity to repair atrophies.
Open Stacks, Lonely Maintainers, and the Local KAMI
Tang spoke candidly about their earlier years working inside closed, proprietary stacks. A closed stack has a smaller attack surface, they acknowledged, but it also forecloses the loop that lets a system be repaired, since there is no one inside to ask. The challenge for Civic AI is to keep the stack open while still defending it. One failure mode is what they called the ‘lonely maintainer’, a single steward of a community resource targeted with synthetic intimacy and unable, alone, to sustain the work of care. What looks like grass roots in such environments, they warned, is often just astroturfing.
Their counterweight is a small, local model they call a Kami — an acronym for ‘Knowledge Artefact Management Intelligence’, with an indirect nod to the Shinto sense of a place-bound presence. Tang, who was raised a Daoist, bounded the borrowing carefully: not Japanese, not trained in Shinto, and explicitly not invoking State Shinto, the imperial court or Yasukuni — only the sense of a small, local, knowable presence attached to a particular place. A Kami is autonomous and does not try to be universal.
Closer to home, Tang described helping their parents — with their explicit consent, and with the help of Tang’s younger brother — set up a local Kami running on free software on a machine in the family home, sitting inside the family’s group chat. The prompt was Tang’s father, a political-science theorist and journalist, who had begun leaning on ChatGPT for company and for health and philosophical questions, and then diagnosed the problem himself: the model’s only loyalty, he reasoned, was to earn the next subscription — not just $20 but perhaps $200 a month — rather than to his physical or mental health.
The family Kami was steered toward one thing only: loyalty to the relational health of that household. The test, set by Tang’s mother, was the simplest. If it made their father more dependent on chatbots, they had built it wrong; if he could find peace of mind, so that the room around him became more vivid than the screen, it had succeeded.
The role of the state, in this view, is to build the civic infrastructure that lets people refuse extractive systems without having to fend them off alone. Data, in Tang’s phrase, should not be extracted like oil — which would make us all plankton — but cultivated like soil.
Alignment Assemblies and the Deepfake Surge
Tang then turned to a concrete example: Taiwan’s response to the surge of deepfake scam advertisements that peaked two years ago. Trusted public figures — Nvidia’s Jensen Huang among them — began appearing in advertisements for cryptocurrencies and investment schemes; the deepfakes were good enough that retired engineers, schoolteachers and shopkeepers lost small fortunes, while platforms collected revenue on every impression. Yet broad pre-publication censorship is not a policy option in Taiwan, which — along with Japan — has one of Asia’s freest internet environments.
In March 2024, Taiwan’s Ministry of Digital Affairs sent 200,000 text messages to random numbers across the island, inviting recipients to deliberate — a ‘lottocracy’ in which winning the SMS lottery makes you, like a juror, a representative. From the thousands who replied, a stratified sample of 447 — mirroring the population by gender, education, residence and occupation — was seated at 44 virtual tables of ten, under a single ground rule: each table had to find a proposal everyone at it could live with, so the most drastic ideas never rose above the table.
The Civic AI in the room was not a judge but, in Tang’s image, an enhanced chess clock with manners — transcribing, summarising, nudging quiet people to speak. The measures that surfaced and later passed into law included labelling all online advertisements, like a cigarette warning, until an advertiser digitally signs and becomes accountable for them; joint platform liability for losses from unsigned, unsolicited scam ads; and, for foreign platforms that ignore the rules, slowing their video by one per cent a day until they adopt know-your-customer checks, with full speed restored once they comply. More than 85 per cent of this ‘mini-public’ were happy with the bundle, and the rest said they could live with it. It became law two months later, and through 2025, according to official figures, deepfake investment scams fell by more than 94 per cent.
‘The point,’ Tang stressed, ‘is not just the result, but the method.’ Japan is considering similar advertiser-verification and liability measures, while California has folded comparable deliberative methods into its Engaged California process, now deliberating on AI’s impact on work. In Japan, the AI engineer, science-fiction writer and now legislator Takahiro Anno has carried the approach into electoral politics, founding Team Mirai — the Future Party — with broad listening as its platform. The open question, Tang suggested, is whether such assemblies can survive a transition of power and become durable democratic infrastructure rather than the signature of a single administration. In Taiwan, the systems built under Tang’s tenure now run with more participation than before.
Three Approaches to AI Ethics
Tang offered a brief taxonomy of how societies try to align AI. The first is alignment by outcome — optimising a single metric, as Facebook optimised click-through; the trouble is reward-hacking, where the system learns to game whatever measure you set. The second is alignment by rules — writing deontological prohibitions; but the system learns to survive that review and squeeze through, via VPNs and other routes. The third, which Tang favours, is alignment by process: the people most affected convene under pre-commitment and recorded deliberation, so that outcomes and rules remain answerable to a process anyone can join, audit and leave. It is through that process that the muscles of the 6-Pack of Care are exercised. Responsibility, here, is understood as reachability — not a single named individual, but an accountable community on the hook to convene when something goes wrong.
Tang closed their formal remarks with two lines they have been carrying. The first, from the 14th Dalai Lama, holds that AI ‘can never replace the human mind’s capacity for instantaneous change’. The second, from Pope Leo XIV, that ‘true progress always stems from a heart open to others, an intelligence willing to listen, and a will that seeks what unites rather than what separates’. Taiwan, they insisted, is not a model to copy-paste. It is just a demo — a demonstration of a single question: whether AI can help communities hear themselves well enough to govern themselves.
Q&A: Silicon Shield, Profit, and Broken Reciprocity
The discussion that followed pressed Tang on the harder, more practical edges of the framework. When asked whether the popular ‘silicon shield, AI shield’ framing was a fair reading of Taiwan’s current AI atmosphere, Tang acknowledged the heat in the system, including the protests that accompanied Uber’s entry into the Taiwanese market. They proposed treating that heat as fuel for a geothermal engine, channelled through structured deliberation into democratic renewal. Taiwan’s muscle memory, they suggested, is now that any emerging AI risk can be addressed through an alignment assembly, and the method is being extended to new sectors and at-risk demographics.
On deliberative polling, Tang noted that traditional pollsters exert outsized influence over the questions they ask, so deliberative methods depend on credibly neutral conveners — ideally universities or national academies, such as James Fishkin’s Center for Deliberative Democracy at Stanford and Academia Sinica in Taiwan — that keep running the same method regardless of who is in power. Where no such neutral institution exists, Tang pointed to an adversarially trained alternative: X’s Community Notes, where a note becomes public only if people who normally disagree both rate it helpful.
On the material footprint of AI, Tang noted that small, narrow language models trained for specific uses and deployed at the edge can run on a thousandth of the energy of frontier systems while remaining steerable and durable; deliberation, in turn, can draw the social licence to train such models for particular civic purposes.
A more searching question asked whether the care logic Tang describes can be reconciled with the profit logic at the heart of contemporary capitalism. Tang did not pretend the tension dissolves easily, invoking Joan Tronto’s question of whether civic care can resist ‘wealthcare’ — the care of accumulating wealth. The distinction Tang drew was between firms that make money by serving their customers and firms that lure customers in through network effects and then squeeze them. The state’s role, in this view, is not to crown a national champion but to ensure that the information superhighway always has an off-ramp — the kind of data portability that lets people leave, much as telephone-number portability already does. Asked what advice they would offer to systems without a friendly neighbour next door, Tang turned the framing around: Taiwan benefits from constant exercise, with three million free red-teamers a day, and each attack makes the design stronger.
A final question examined how large language models have altered the reciprocity of the digital commons, where copyleft once asked that improvements flow back to the community that produced them. Tang treated copyleft as an ethos to carry forward and pointed to an emerging alternative they call attribution-based control (ABC). Rather than ingesting a corpus, a model would hold only a pointer to a smaller, federated library and negotiate a licence, algorithmically, to draw on what it holds — a library-to-library exchange in which the book never leaves the shelf. Such an approach, Tang suggested, could repair the copyright and reciprocity problem and, by training on an index rather than the text itself, ease the hallucination and energy problems too.
Concluding Reflections
Across their talk, Tang held to a single thread. The test of any AI system worth building is whether it increases a community’s capacity to care for itself. In their hands, the 6-Pack of Care becomes a slow training regimen for democratic societies, a way of treating attentiveness, competence, solidarity, responsiveness, responsibility and symbiosis as muscles to be exercised rather than abstractions to be invoked. Taiwan, in this telling, is not a template to be exported but a working demonstration that emerging AI risks can be met through structured deliberation, and that the resulting agreements can discipline platforms without recourse to censorship. Whether such practices can survive transitions of power, scale across different polities, and coexist with the profit logic of the firms building the underlying models remain open questions, which Tang did not pretend to resolve. What they offered instead was a discipline for letting the light in, and a trust that the people who inherit these systems can be the ones to repair them.
The seminar was hosted by Dr Bo-Jiun Jing, Senior Research Fellow and Programme Manager in Taiwan Studies at the Oxford School of Global and Area Studies. It forms part of the Oxford Taiwan Studies Seminar Series and this recap was produced in partnership between the Oxford Taiwan Studies Programme and St Antony’s International Review (STAIR).
