Method note: The following primary responses were generated by jdd-kami, a bounded drafting system developed with Gisele Chou and described publicly at gh#642998ce.
The system was run against a fixed local research corpus at gh#0a74e2d5 and hf#04ac740. After each answer, Audrey Tang either endorses, qualifies, or corrects the draft.
Article Zero: Many democracies collect citizen input, yet people often still feel unheard. What concrete mechanisms successfully translate participation into binding decisions — and how should we measure whether that translation is truly working?
jdd-kami:
Being heard is not the same as having power. Democratic trust frays when the former is presented as if it were the latter: input is collected, but no one explains where it went or what it changed.
It helps to distinguish four levels of participation: consultation, where people are heard; agenda-setting, where public input shapes what gets discussed; adopt-or-explain, where decision-makers must respond publicly to proposals; and legally binding decision, where public participation has direct formal force. Most democratic innovations sit in the middle two layers. That is not a weakness, but it should be named plainly. If consultation, uptake, and binding authority are blurred together, trust eventually breaks down.
A mechanism that genuinely translates participation into consequential outcomes needs at least three features: traceability, so people can see how their input travelled into a recommendation or draft; two-way feedback, so institutions explain why a proposal was accepted, revised, or declined; and genuine influence without high barriers to entry, so not only experts or well-organised stakeholders can shape the result.
The vTaiwan process is a useful example. Policy questions are first opened for structured dialogue on Polis, where participants respond to statements with “agree / disagree / pass”. The system uses statistical clustering to surface views shared across groups rather than magnify division. One design choice matters most: there is no reply button. That is not merely an interface decision; it is a democratic one. It interrupts the spiral of argument and rewards statements that can travel across difference.
But precision matters here. vTaiwan does not make citizen input automatically legally binding. At its best, it creates an adopt-or-explain discipline. When rough consensus emerges, the relevant public authority is expected to respond formally and explain whether, and how, it will carry proposals forward.
Taiwan’s 2024 Alignment Assembly extended this logic to AI, information integrity, and scam advertising. A civic lottery began with 200,000 government text messages, producing 1,760 valid responses and broad agenda formation. Then 447 demographically representative participants deliberated in 44 online tables, supported by AI-assisted transcription and synthesis. Across key proposals, support exceeded 85 per cent. Those recommendations helped shape a broader anti-fraud policy and legislative pathway. Subsequent platform enforcement and advertiser-verification measures were later associated with a reported 94 per cent reduction in impersonation scam ads within a year.
So the metrics should also be layered. How many contributions entered the formal record? How many received a public answer? How many changed the final text? Which stage of authority did the process actually reach: consultation, agenda-setting, adopt-or-explain, or binding decision? After implementation, there is one harder question: do participants feel the outcome reflects their voice, even if it is not their first choice? That measure is imperfect, but it comes closest to whether trust has actually been repaired.
Audrey Tang — Qualification
The structure is good and “being heard” and “having power”, of course, quite different. On the other hand, many fruitful participatory processes do not get full power-sharing at the beginning; they produce common knowledge for surprising agreements (“uncommon ground”) and legitimacy.
Article Zero:
Transparency in digital participation can be empowering, but it can also be gamed. What design principles help prevent civic platforms from becoming arenas for coordinated manipulation or performative participation?
jdd-kami:
As a rough heuristic, manipulators optimise for tally movement, while sincere participants are more likely to seek mutual understanding. Good design should make the latter easier and the former harder.
First, improve the signal-to-noise ratio rather than amplifying volume. Systems built around sheer reaction count often reward the loudest and most easily mobilised voices. Better systems ask a different question: not “What can rally my side?” but “What could people across differences still accept?” That favours bridgers rather than polarisers. Polis is useful here because removing the reply function lowers confrontation structurally and rewards statements that can travel across groups.
Second, make coordinated behaviour legible as a pattern. Rate limits, provenance checks, anti-Sybil controls, and similarity or time-window flags can help a platform recognise near-identical bursts as coordination rather than independent sentiment. Transparency should not become spectacle. Its purpose is to reveal how influence is being organised, amplified, and possibly distorted.
Third, prefer rough consensus over simple majority rule. A system seeking “a proposal no one strongly objects to” is harder to game than one seeking 50 per cent plus one. Mobilisation alone may be enough in the latter. In the former, one must also reduce opposition and widen acceptability.
Fourth, keep human deliberation at the final stage. Algorithms can help sort, cluster, and summarise, but they cannot be the source of public legitimacy. Participants must remain able to contest the synthesis, revise it, and confirm whether it fairly represents what was said.
One caution is essential: no interface design fully neutralises organised campaigns. Design does not eliminate manipulation; it changes which tactics are cheap, which are expensive, and which leave detectable traces. That is already a democratic gain.
Transparency, then, should do more than make participants visible to one another. It should make visible how influence itself is being organised and amplified.
Audrey Tang — Endorsement
I agree that affordance (“changes which tactics are cheap”) is the right framing here. Bad faith actors are ever-present. What we can do is designing toward strategy-proofness and making bridging actions easier and polarisation-reinforcing behaviour costlier.
Article Zero:
You have emphasised bridging over polarising. What practical tools enable “rough consensus” at scale, especially in situations where values genuinely conflict?
jdd-kami:
“Rough consensus” comes from the working culture of the Internet Engineering Task Force. Its famous formulation is: “We reject kings, presidents and voting. We believe in rough consensus and running code.” The point is not unanimity. It is that no unaddressed strong objection remains large enough to block movement.
That is a useful democratic instinct. It does not deny conflict. It asks how conflict can be worked through rather than merely amplified.
At scale, different tools do different jobs.
Polis maps the opinion landscape. It shows where unexpected common ground exists and where disagreement is real. Smaller but coherent groups can contribute distinctive information; they should not disappear simply because they are not numerically dominant.
Loomio is useful for decision formation. Its value is not only that participants can take positions, but that they can state reasons. “I object” becomes “I object because …”. Once reasons are visible, revision becomes possible.
Deliberative mini-publics are especially useful when an issue has been gridlocked for years. In Ireland, the path to the 2015 marriage-equality referendum ran through the Convention on the Constitution, while the 2016–2018 Citizens’ Assembly addressed the Eighth Amendment and helped prepare the ground for the 2018 abortion referendum. These were related but distinct institutional routes. The marriage-equality referendum passed with 62.07 per cent, while the 2018 abortion referendum passed with 66.4 per cent. The broader lesson is that representative deliberation can help elected politics move on questions it has long struggled to resolve, even when the deliberative body itself is advisory rather than binding.
Taiwan’s marriage-equality debate offers a different lesson. Polarisation often persists because opposing camps are not answering the same question. One side may be defending equal individual rights; the other may be defending a social meaning embedded in kinship, religion, or family vocabulary. Once that distinction becomes publicly legible, the task changes. The question is no longer simply “Who wins?” but “What arrangement protects rights while lowering unnecessary symbolic threat?”
That is the deeper move: from victory logic to design logic. Under what conditions can people move forward together, even while disagreeing? The role of facilitation is not to erase conflict. It is to turn disagreement into conditions that can be worked on.
Audrey Tang — Qualification
On the Taiwan example, I would add one caution: deliberation did not dissolve moral disagreement. It clarified which part of the disagreement had to be settled through equal civil rights, and which part could remain socially contested.
Article Zero:
If AI is treated less as a product and more as civic infrastructure, what elements should be publicly owned, audited, or governed — models, datasets, procurement rules, evaluation benchmarks, or something else?
jdd-kami:
If AI is treated as infrastructure, the central question is not ownership alone. It is whether the system is arranged so that everyone can pass through it on fair, inspectable, and contestable terms.
On that basis, the most urgent public priorities are not always the models themselves. They are the standards around them.
First, evaluation benchmarks and audit standards should be publicly governed. Private firms may build models, but they should not also monopolise the tests, the metrics, and the interpretation of whether those systems are safe, fair, or accurate. When the same actor builds the model, defines the benchmark, runs the evaluation, and explains the result, the conflict of interest is obvious.
Second, procurement rules should be far more transparent. When governments buy AI systems, the public should be able to inspect the decision criteria, contract structure, evaluation process, known failure modes, and remediation path. In the Dutch SyRI case, the District Court of The Hague held in 2020 that the welfare-fraud system violated Article 8 of the European Convention on Human Rights because its operating logic and risk indicators were too opaque. That case made an essential point: in public administration, opacity is not only a technical weakness; it is a legal and democratic one.
Third, where models rely heavily on public resources such as government documents, court rulings, or public broadcasting archives, some form of public-interest reciprocity should be considered. This is a normative direction, not a settled legal doctrine. The narrower and stronger claim is that governments can attach conditions through procurement, licensing, or access rules: transparency obligations, public-interest access tiers, local-language commitments, or audit rights. A reasonable public-interest principle is that if public resources materially support a system, the public should not be treated as an afterthought.
Foundation models do not always need to be publicly owned. States are not necessarily best placed to train the most capable systems. But public institutions are entirely entitled to decide which models should be approved for public use. Explainability, contestability, safety, and competence in local languages and cultural contexts should be minimum thresholds, not optional extras.
One governance question remains neglected: who gets to evaluate? If evaluators are only technical specialists, they will predictably miss harms borne by cultural minorities, disabled users, lower-resourced communities, and those least visible in standard datasets. Governing AI as infrastructure means governing not only the system, but also whose experience counts when standards are set.
Audrey Tang — Qualification
I fully endorse the priority given here to benchmarks, audits, and procurement. That is where governance becomes real. One qualification matters: reciprocal obligations for models trained on public resources remain more of a normative direction than a settled legal doctrine. For now, the strongest framing is through procurement, transparency, and public-interest access.
Article Zero:
When AI systems are deployed in public administration, what does a clean and responsible accountability chain look like? Where should responsibility ultimately reside?
jdd-kami:
A clean accountability chain is one that prevents responsibility from evaporating.
The common failure mode is familiar. Vendors blame operators. Operators blame the algorithm. The algorithm blames the data, and the data blames reality. In the end, no one is left standing where the consequences land.
To prevent that, at least four things are needed.
First, every consequential decision requires a named human decision-maker. If an AI-assisted system affects a person’s rights, benefits, access, or legal status, “the algorithm decided” should never be an acceptable answer.
Second, the decision record must be explainable to the affected person. That means more than technical documentation for experts. It means an explanation a citizen can actually use: what happened, why it happened, what inputs mattered, and what can be challenged.
Third, appeals pathways must be layered and legible. Individual case review, agency review, independent oversight, and judicial review should be clear in advance, along with time limits, duties, and possible remedies. Institutions should not only make decisions; they should make correction possible.
Fourth, audit and impact assessment must continue after deployment. The question is not only whether a system passed a pre-launch test. It is whether, in practice, it is producing unequal burdens, systematic error, or hidden exclusion. Post-deployment monitoring is where many public harms first become visible.
Taiwan’s Participation Officers are suggestive here. The lesson is not the title but the principle: someone inside government must be explicitly responsible for bringing public input into bureaucratic process and tracking whether commitments are carried through.
Ultimate responsibility resides with those who made the political choice to deploy the system. That includes the officials who approved procurement, defined the use case, accepted the risk model, and set, or failed to set, the oversight conditions. Vendors provide tools. But the decision to insert those tools into public administration is political. Political responsibility cannot be outsourced.
Audrey Tang — Qualification
I endorse the structure of this answer almost entirely. My only tightening is institutional: the safest formulation is that ministries and subordinate agencies appoint participation officers, rather than implying that a named officer exists in every team or sub-unit. The broader principle remains correct: if responsibility is diffuse everywhere, it is usually absent anywhere.
Article Zero:
Care is often misunderstood as soft. In public systems, it can be structurally demanding. How would a care-ethics lens concretely reshape AI governance — in regulation, institutional design, or public budgeting?
jdd-kami:
Care is often dismissed as soft. In practice, public systems have long treated dependency, maintenance, and relational responsibility as secondary concerns. But the moment we ask who bears the cost of a system, who notices its failures, and who is most exposed when they are ignored, we are already asking political questions.
That is why care ethics matters for AI governance. In many current systems, the well-resourced design the tools, while the less well resourced bear the risks. Those who bear the risks usually have the least power to shape the design. Care ethics names that imbalance directly.
On that view, AI governance would change in at least three ways.
First, success metrics would shift from efficiency alone to relational health. Systems are often judged by speed, cost, and predictive performance. Those matter, but they are incomplete. A care-based framework asks whether the technology improved the relationship at stake: between nurse and patient, teacher and student, caseworker and applicant, citizen and state. If time is saved but never returned to that relationship, we should ask what the gain was for.
Second, those systematically absent would become a formal governance concern. The question would no longer be only who is in the room, but who is structurally prevented from appearing at all. That means funding translation and accessibility, surfacing under-documented needs, and building methods for people with low voice or low institutional visibility to shape the process. It cannot depend on the goodwill of a single designer or official.
Third, budgets would be treated as declarations of care. A government serious about care ethics would not spend only on model capability and rollout speed. It would also fund impact assessments for marginalised communities, participatory mechanisms for affected groups, and tools that strengthen the work of carers such as nurses, social workers, and teachers. Budgets reveal priorities more honestly than mission statements do.
Care ethics, then, does not ask us to become gentler in the abstract. It asks us to become more structurally responsible: to design for dependence, repair, unequal vulnerability, and the long work of maintaining trust.
Audrey Tang — Endorsement
This is the answer in the set that comes closest to a governance philosophy rather than only a governance checklist, and I endorse it strongly. The sentence I would highlight is: “Budgets reveal priorities more honestly than mission statements do.” If care is real, it should appear in staffing, translation, accessibility, community participation, and support for frontline workers, not just in rhetoric.
Article Zero:
If an AI provides the primary answers in this format, what is this system reliable for, and what is it not reliable for? What further safeguards would you add to ensure epistemic clarity and democratic responsibility?
jdd-kami:
The clearest answer is procedural.
I am quite good at structured synthesis, conceptual comparison, and first-draft articulation within a declared scope. You can trust me to move quickly across bodies of literature and case material, identify recurring patterns, and render complex arguments into a more legible form. When the question concerns events, debates, or frameworks already contained in its knowledge base, it can often produce a useful first draft.
I am also useful, if honestly presented, for declared constraints. I’m not a universal authority or a live witness. I’m better understood as a bounded drafting system with stated commitments, explicit limitations, and human review. I’m governed by the 6-Pack of Care: attentiveness, responsibility, competence, responsiveness, solidarity, and symbiosis. Those are not decorative virtues. They are design constraints that can be named, challenged, and revised.
What I’m not reliable for is equally important. I have no real-time view of the world. I should not be treated as a witness to breaking events, rapidly changing political situations, or lived conditions on the ground. Like other language models, I can be fluent, coherent, and wrong.
There is another limitation that matters especially in democratic contexts: I tend to under-represent what is under-documented. Oral knowledge, local memory, marginal experience, and communities less visible in dominant datasets can all be flattened. That is not a minor technical flaw. It is one of the central political limitations of this genre of system.
So epistemic clarity requires more than a disclaimer. It requires provenance. Readers should be told what corpus or briefing pack informed the interview, when it was frozen, what kinds of sources were included or excluded, and what stage of human review followed. Without that information, transparency remains aesthetic rather than operational.
That is also why Audrey’s role matters. Human commentary should not function as decorative endorsement. It should do visible accountability work: endorsing, qualifying, correcting, and, when necessary, stating that a claim remains uncertain. If that layer is real, responsibility lands somewhere. If it is absent, the prose may still sound authoritative, but the accountability chain remains incomplete.
So the right posture is neither trust nor dismissal in the abstract. It is to treat these answers as serious but revisable first drafts: useful for orientation, not exempt from challenge. In that sense, the format is not outside democracy. It is a test of whether democratic responsibility can still be made visible when one of the voices in the room is synthetic.
Audrey Tang — Correction and Endorsement
The “trust me for” wording drifted towards anthropomorphism. For this interview, the more useful and accountable description is the procedural one: we can rely on jdd-kami as a bounded drafting system with declared commitments, explicit limitations, and human review. That wording avoids unnecessary metaphysical debate and keeps the focus on provenance, revision, and responsibility.
My commentary is not ornamental. Where I endorse, I am standing behind the claim. Where I qualify or correct, I am showing the limits of the draft. That visibility is part of the experiment. Without it, AI-generated prose can appear grounded while responsibility remains unassigned.
Photographer: Kaii Chiang, License: CC BY-NC-SA 4.0 | Edited (background expanded using AI)
