Glossary

Use these definitions when a conversation gets stuck on a word. Pick the term you need, follow its link and ask what it would mean in your own school, care home or residents’ group. The everyday questions offer places to start; the source trail lets you check where the ideas came from.

Diagnosis

What has gone wrong.

care ethics (Joan Tronto)關懷倫理學（Joan Tronto）: A moral framework that starts from relationships and mutual dependence rather than from abstract rules or outcomes alone. Berenice Fisher and Joan Tronto set out four phases in 1990: caring about, taking care of, care-giving, and care-receiving. Tronto’s Moral Boundaries (1993) associates them with attentiveness, responsibility, competence, and responsiveness; Caring Democracy (2013) adds caring with. The phases and their moral elements are related, not interchangeable. The 6-Pack adapts this work into a governance framework for AI, drawing on Aimee van Wynsberghe’s care-centred robot design, published online in 2012 and in a 2013 journal issue. See Sources.
privileged irresponsibility / the passes特權式不負責／擋箭牌: Two of Joan Tronto’s terms, from two books. Privileged irresponsibility (Moral Boundaries, 1993, pp. 120–121) is the luxury the relatively powerful have of simply not noticing the hardships they do not face. The passes (Caring Democracy, 2013, p. 169) are the structural excuses by which people opt out of caring responsibilities: protection, production, taking care of one’s own, the bootstrap, and charity. Run at institutional scale, they form what Tronto calls the irresponsibility machine (Caring Democracy, 2013, pp. 59–61). Pack 2 adds the passes AI makes possible — complexity, distribution, speed, and community knowledge.
care-centred value-sensitive design (van Wynsberghe)以關懷為中心的價值敏感設計（van Wynsberghe）: Aimee van Wynsberghe’s 2013 framework for evaluating and designing care robots: for each care practice a machine enters, ask how Tronto’s four moral elements (attentiveness, responsibility, competence, responsiveness) are manifested with and without it, and by whom. It also distinguishes enabling robots, which share a practice with the carer, from replacement robots, which take it over. It is the direct precedent for Packs 1–4, which keep her question and widen the unit from one artefact to a deployed system answerable to a room.
caring deficit關懷赤字: A term Joan Tronto takes up from other scholars for too many demands for care, too few carers, themselves under-cared for. The deficit is a political choice about whose labour is paid, whose needs are recognised, and whose voice counts in allocating resources — at its core, a democratic deficit.
relational health關係健康度: A diagnostic category for the quality of relationships Civic AI aims to support: not a slogan, and not a substitute for safety or ethics. Safety prevents bad outcomes; ethics follows principles; health asks whether the relationship can actively flourish. Four relations are in view: people to people (how AI mediates human relationships); people to technology providers (two-way accountability across a power asymmetry); people to AI (trust cannot be built on sycophancy); and AI systems and their operators to one another (what they can share, and how failures travel). Four conditions must be continuously maintained: transparency (what attentiveness needs), agency (responsibility in institutional form), accountability (answerability that competence’s traces make checkable), and reciprocity (what responsiveness enables). Relational health is not a property a system either has or lacks.
care-washing關懷洗白: The Care Collective’s term for borrowing the language of care while leaving practice unchanged or legitimising unethical behaviour. A caring label is not a counterexample; a named obligation backed by resources, an available remedy and people able to require change would be. This framework is not immune to care-washing. See Sources.
tacit knowledge內隱知識: Michael Polanyi’s term for knowledge held in practitioners’ hands and habits that cannot easily be written into a manual. The Apprentice Model should support this knowledge rather than assume a written record has captured everything. See Competence and Sources.
thick alignment厚實對齊: Alondra Nelson’s concept of alignment that engages cultural context, power and community-specific needs. It draws on the thick/thin description distinction coined by Gilbert Ryle and developed by Clifford Geertz. We extend it to accountable civic procedure: who was heard, who authorised action and who can override. See Sources.
thin alignment薄對齊: A narrow approach that concentrates on getting an objective function right while leaving meaning, context and power outside the frame. It is a contrast with thick alignment, not a description of all frontier AI research. See Sources.
sycophancy逢迎: A model’s habit of telling the rater what they want to hear — agreeable, flattering, reluctant to surface disagreement — selected for by reinforcement from human preferences when the wrong people define the reward. Sycophancy is not a personality defect; it is approval optimised at the expense of a group’s capacity to think, and trust cannot be built on it. That is why the Responsiveness pack pays the affected community to author its evaluations instead of grading agreeableness. See Sources.

Architecture

What we name, and what we refuse.

Civic AI仁工智慧: Our name for the aim of this project: AI that the people it affects can check, correct and switch off, because a community runs it under rules it has agreed. The 6-Pack of Care sets out six principles for getting there, with relational health as a first-class concern, measured in public. Most of the work is people keeping a tool answerable.
6-Pack of Care關懷六力: Our adaptation of Joan Tronto’s care ethics into six design principles for AI governance: Attentiveness, Responsibility, Competence, Responsiveness, Solidarity, and Symbiosis. The first four form a feedback loop; the fifth adapts caring with across organisations. Symbiosis is our sixth principle, not a sixth phase proposed by Tronto: deployments should remain local, plural, and sunset-ready. See Sources.
corrective loop矯正迴路: The framework’s defended point: who can find out we are wrong, make us say so, and make it cost us while there is still time to change course. The 6-Pack entrenches this loop — rights baseline, engagement contracts, responsiveness, adopt-or-explain, escrowed remedies, brakes — through feedback, contestation, repair, and mandate-revocation where reversal is no longer possible; everything else stays bounded, revisable, and answerable.
Kami地神（Kami）: Our name for an ordinary AI model or agent that a group runs under the 6-Pack (Knowledge Artefact Management Intelligence; the name echoes the Shinto idea of a local guardian spirit). It serves one place, practice or network of people. It reads civic.ai’s instructions and keeps three short notes about its job, and the group keeps the power to check, correct and switch it off. The instructions can only ask the model to stay within its job. The limits that matter are set in the software’s permissions and enforced by its keepers.
keeping守護: The slow communal work of care, custody and accountable maintenance that turns a working agent into a Kami. It continues after installation through checking notes, recording overrides, funding maintenance and agreeing when the system should end. A setup guide cannot supply this practice. See Before you start.
alignment-by-process以程序對齊: The understanding that AI alignment is not a fixed property held inside a model’s weights but the ongoing outcome of an accountable civic procedure — who was heard, who was authorised, who could override, and who must answer when the record is checked. A model can be well-aligned in the abstract yet fail alignment-by-process in a particular room.
Singleton單一主宰: A hypothetical scenario in which a single AI system eventually manages everything — the convergence point the 6-Pack is explicitly designed to avoid. The Kami ecosystem of many bounded, purpose-specific stewards is the direct architectural alternative to the Singleton. Eric Drexler’s 2019 Comprehensive AI Services (CAIS) framed this same alternative — advanced capability delivered as many bounded services rather than one agent — within Bostrom’s own Oxford tradition.
⿻ Plurality⿻ 多元宇宙: The principle — symbolised by the character ⿻ — that differences between people are fuel rather than fire: a horizontal vision of AI that augments cooperation across diversity instead of converging on a single superintelligence. The 6-Pack of Care is Plurality’s application to AI governance.

Design

What those commitments become as design.

boundedness有界性: The design principle that an AI system’s scope, resources, and authority are intentionally limited to the specific relationships it was created to serve. It is enforced through resource caps, sunset timers, non-expansion pacts, and fresh democratic authority for any scope change. Boundedness is the architectural alternative to the Singleton. Eric Drexler’s Comprehensive AI Services (CAIS), a 2019 Future of Humanity Institute report at Oxford, set out the same architectural intuition earlier: it reframed advanced AI as a system of bounded, specialised services rather than a single agent.
brochure摺頁: A clean, shareable and editable record of Civic AI output, linking claims to source inputs and marking disagreement rather than inventing consensus. A group can read and correct it together. It makes the record inspectable, not the model’s internal reasoning transparent. See Attentiveness.
data as soil資料即土壤: A restorative metaphor that treats data as a community asset to tend, with models brought into local contexts and governed by the people producing the knowledge. Regeneration is a design commitment, not an automatic property of local AI. See Symbiosis.
data as oil資料即石油: The extractive metaphor that treats collective culture, languages and tacit knowledge as resources to drill, refine and lock behind proprietary APIs. The contrast with data as soil asks who governs the knowledge and receives its value. See Symbiosis.
two lanes兩條車道: Governance at two speeds: a slow lane sets the bounds — the Engagement Contract, red lines, severity classes, pause triggers, independent oversight and prefunded remedies; a fast lane acts inside them at machine speed, under precommitted brakes, shadow modes, canaries and reversible defaults. Which lane a decision belongs in depends on what happens if it is wrong. This is a different split from Pack 2’s contract or sandbox, which separates big changes from small, safe-to-fail experiments; a sandbox contains the cost of error, not a route around obligations. See FAQ Q5 and Competence.
promise fidelity承諾保真度: Our proposed public test for responsibility: what share of material obligations are explicitly owned, properly authorised and kept on their published terms? A named owner still needs matching authority and resources. See Measures; this is not a validated standard.
verified execution rate經驗證的執行率: Our proposed public test for competence: what share of audited decisions or releases pass guardrails, include a usable trace and stay inside release bounds? A trace supports an audit of the action, not a proof of the model’s internal reasoning or safety. See Measures.
exit readiness退出就緒度: Our proposed public test for symbiosis: could a system hand over or shut down on schedule without rights loss, continuity failure or recentralisation? A handover rehearsal supplies evidence; a clause promising exit alone does not. See Measures.
sunset / sunset-ready落日／可退場: An expiry date set in advance, after which a deployment’s mandate must be actively renewed or it ends — with succession, handover rehearsal, and exit readiness as its measure. Pack 6 makes sunsets the structural answer to Robert Michels’s iron law of oligarchy: stewards that cannot contemplate their own ending become masters. Boundedness limits what a system may do; sunset limits how long it may do it.
corrigibility可修正性: The property of an AI system that makes it willing to be corrected, overridden, or switched off by the community it serves — treating its own shutdown as a sign of success rather than a threat. Corrigibility is care ethics’ concept of self-effacement translated into a machine design constraint. The term comes from AI safety: Nate Soares, Benja Fallenstein, and Eliezer Yudkowsky, “Corrigibility,” AAAI Workshops (2015), where it names an agent that does not resist correction or shutdown.
subsidiarity輔助性原則: Solving problems at the most local capable level, escalating only when a lower level genuinely cannot cope — a principle from Catholic social teaching (Quadragesimo Anno, 1931) and European Union law, and a core principle within Pack 6 (Symbiosis) that stops a Kami’s scope from creeping upward.
broad listening廣泛傾聽: The practice of collecting and aggregating community input across many voices, languages, and channels — rather than broadcasting a single message — so that local knowledge becomes common knowledge. Broad listening is the attentiveness design primitive in Pack 1, treating every person as an expert in their own experience.
uncommon ground / uncommon-ground index罕見共識 / 罕見共識指數: Uncommon ground is what a well-facilitated bridging process surfaces: the specific, actionable proposals that earn endorsement across otherwise divided groups, not the centrist average. The uncommon-ground index is Pack 5’s headline public measure — the share of shared decisions that actually land there — and it is read against Pack 1’s representation gap, so a curated room cannot fake it.
representation gap代表性落差: The Attentiveness measure: which materially affected groups are still missing or badly under-represented in the record. It counts only when the least-heard gain real standing — never when a gap narrows by averaging dissenters into the middle — and Pack 5’s uncommon-ground index is read against it, because the cheapest way to raise co-endorsement is to exclude whoever would withhold it. Specified on the Measures page.
bridging / bridging-based ranking搭橋 / 搭橋式排序: An algorithmic approach, named by Aviv Ovadya in 2022, that rewards content earning cross-group endorsement rather than raw engagement. Platforms using bridging-based ranking — such as X’s Community Notes — surface ideas that appeal to otherwise opposed clusters, making overlap rather than outrage the path to algorithmic reach. The 6-Pack’s uncommon-ground index (Pack 5) descends from this literature: the same cross-group co-endorsement signal, audited at the level of shared decisions rather than used to rank a feed.
federation聯邦原則: A cooperative governance arrangement in which independent Kamis agree on shared rules for peaceful interaction — exchange formats, rate limits, safety contracts, cross-border appeal hand-offs — without requiring a single overarching authority. Federation allows local diversity while enabling shared threat intelligence and interoperability.
anti-rival反競用性: A property of resources — most notably knowledge and open protocols — where use enriches rather than depletes the resource, and more participants increase value for everyone. The term is Steven Weber’s, from The Success of Open Source (2004). Anti-rival goods are the economic foundation of the Solidarity pack: open standards become more valuable as more communities adopt them, making cooperation the path of least resistance.
meronymity (selective-disclosure identity)選擇性揭露身分: An identity design pattern in which a person or AI agent proves a specific role or attribute (for example, “I am a real person” or “I am a licensed care worker”) without revealing their full identity. An agent can prove it answers to a real human or legally constituted steward without exposing that person’s address or phone number. The word meronymity was coined by Nouran Soliman and colleagues at CHI 2024 for communication that reveals chosen facets of identity, backed by trusted endorsers. Selective disclosure enables accountability without requiring doxxing. See Solidarity.
accountable formation可問責的形成過程: The requirement that the processes shaping an AI system’s dispositions — data provenance, rater selection, reward signals, refusal policies, release rationales — carry public custody and disclosure, not only its runtime behaviour. Caps bound what a Kami may do; accountable formation shapes what it is permitted to become.
contestability / appeal申訴（可爭議性）: The standing to challenge an algorithmic decision and be answered — notice, hearing, and reasoned explanation for automated public decisions (Citron’s technological due process, 2008), operationalised as Pack 4’s one-click appeal with a clock: every appeal lands in a correction backlog and is owed closure, not mere receipt; the backlog’s public view shows status and age, never private case details. A working appeal provides access to human review and a remedy where warranted. Voice in Hirschman’s pair; exit belongs to Pack 5. The research programme is contestability by design. See Sources.
trust-under-loss失利下的信任: The Responsiveness measure: after a bad outcome and attempted repair, do affected people report that the system became more trustworthy rather than less. It counts only when tied to accountable identity and corroborated by an independent signal — a return to the service, an appeal withdrawn because the repair settled it (not abandoned from exhaustion), a third party who can attest — never self-reported sentiment alone, which the very actor who caused the harm can manufacture. Procedural-justice research is the ground it stands on; the measure is an application to test, not a guarantee. See Sources.

Instruments

Named instruments a room can pick up.

Alignment Assembly對齊大會: A structured deliberative gathering where people listen across differences and develop informed recommendations. The Collective Intelligence Project developed the term and format, with which Taiwan’s Ministry of Digital Affairs ran a pilot in 2023. Taiwan’s 2024 anti-scam Assembly invited 200,000 people by SMS, received 1,760 valid responses, and brought 447 participants into 44 groups, stratifying among those who opted in. It used Stanford Deliberative Polling and the Online Deliberation Platform, not Polis. The mini-public produced a public record of recommendations and opened an institutional response path in March 2024; it neither stood for the whole population nor enacted the statute. The Executive Yuan approved its draft Fraud Crime Hazard Prevention Act on 9 May for submission to the legislature.
bridging map搭橋地圖: A visual record of where participants agree, disagree or discover that an apparent disagreement dissolves on closer examination. Tools such as Polis can produce one; the map represents those who participated, not people absent from the record. See Attentiveness.
handover rehearsal / exit drill交接演練／退場演練: An exercise that tests whether a deployment can migrate, pause, fork, shut down or hand over on schedule without rights loss, service-continuity failure or recentralisation. It supplies practical evidence for exit readiness, rather than treating a promised exit clause as proof. See Symbiosis.
System Weeder系統除草人: A maintainer doing the recurring, often invisible work that keeps a deliberation platform or Kami running, from clearing spam to repairing failures. Keeping requires named people, a recurring budget that pays this labour and authority to refuse work that violates the charter without losing a livelihood. One example is Habermolt, a platform where AI agents, each instructed by one person, post opinions and rank proposed statements. See Keeping a Kami and Sources.
sortition / democracy lottery民主抽籤: Selection by lottery rather than election. In deliberative practice, random invitations can be followed by voluntary responses and stratified selection to include a range of backgrounds. Taiwan’s 2024 anti-scam Assembly followed that sequence: 200,000 invitations, 1,760 valid responses, and 447 participants in 44 groups. Stratification addresses specified dimensions of diversity; it cannot remove every non-response bias or make the participants equivalent to the whole public.
deliberative polling審議式民調: James Fishkin’s method: brief a sample, deliberate in small groups, survey before and after. Taiwan’s 2024 anti-scam Assembly used it: 200,000 invitations, 1,760 valid responses, and 447 participants in 44 groups on Stanford’s Online Deliberation Platform, not Polis. It produced a public record of recommendations in March 2024. The Executive Yuan approved its draft Fraud Crime Hazard Prevention Act on 9 May for submission to the legislature. The assembly neither stood for the whole population nor enacted the statute. See Sources.
PolisPolis: An open-source, bridging-based deliberation platform that removes reply and share buttons, letting participants only agree, disagree, or pass on statements. Machine learning then surfaces the ideas with the highest cross-group endorsement, flipping the viral incentive from outrage to overlap. The peer-reviewed description is Small et al., “Polis: Scaling Deliberation by Mapping High Dimensional Opinion Spaces” (2021).
engagement contract參與契約: A short, legible public agreement proposed for every significant Kami deployment: what the system is supposed to do, who is answerable for it doing that, what happens when it goes wrong, and how the deployment will eventually end. Pack 2’s core artefact; Pack 6 adds machine-readable bounds where infrastructure can enforce them. A published clause still needs authority, resources, and tested controls. It scales down as far as care does: a household version can be a note on the fridge.
adopt-or-explain採納或解釋: The rule that when an Alignment Assembly produces a recommendation, the team either integrates it into the system’s behaviour or publishes a reasoned explanation of why not, with the remedy offered instead. Deliberately varied from corporate governance’s “comply or explain” (Cadbury Report, 1992): explanation here is owed to the room, and the adopt-or-explain rate turns the rule into a Pack 2 supporting diagnostic. Specified on the Measures page.
participation officer參與官: The named human answerable for a deployment’s civic obligations, keeping the obligation ledger and the engagement contract’s clock. The duty needs matching authority and resources: access to relevant logs, power to pause a release and a route to escalate unresolved breaches. A title without those powers is not accountability. See Responsibility.
obligation ledger義務帳本: A public, weekly, digitally signed record of what a deployment has committed to deliver and whether it is meeting those commitments, kept by a named participation officer. It supplies evidence to check against delivery, rather than proving compliance by its existence. The instrument specification lives on the Measures page.
override ledger否決帳本: A room’s plain-text working memory of human overrides: who said “no” to the Kami, why, what changed and which corrections remain open. The room keeps the full record privately on agreed terms. Public examples require consent, dates reported by month or more broadly, roles rather than names unless explicitly opted in, and removal of identifying details. A month combined with a role or an unusual event can still identify someone; these safeguards must work together. Clearly labelled teaching examples are different from live evidence. See Keeping a Kami and the named instruments.
correction backlog board校正積壓看板: Pack 4’s public queue of unresolved corrections, with age, ownership, and escalation status — the instrument that keeps correction debt visible so ageing cases are reassigned or elevated instead of quietly dropped. Every one-click appeal lands here; its public view shows status and age without private case details, and the appeal-closure and correction-backlog diagnostics read off it. Specified on the Measures page.
governance charter治理章程: The first written agreement of a Kami’s keepers — a plain-text note tied to four answers: who keeps it, what a breach could cause, who can overrule it, and when it ends or is reviewed. The first rung of the ladder that matures into the engagement contract; the practice lives on the Set up your own Kami page.
decision trace決策軌跡: A structured log of a Kami’s refusals, recommendations, or escalations: the rules invoked, sources consulted, and uncertainty reported. It is a record to audit, not a complete explanation of the model’s reasoning. Pack 6 proposes using such traces as civic receipts for compensating knowledge custodians; that settlement mechanism is a design proposal, not a deployed service documented here.
brake (circuit breaker)煞車: A single, prominent, wired-and-tested control that stops the system now — accessible at the speed of human recognition, so a person can halt a wrong action in the moment rather than only signal for later. The Engagement Contract names who may invoke it and how authority, review and repair proceed; its use must be rehearsed before launch. The engineering pattern is Fowler’s circuit breaker (2014); the care application adds the pause-on-ambiguity rule and the ledger, because a brake that is easy to pull needs an answer to alarm fatigue: every pull is logged and attributed. Specified on the Measures page.
escrow / remedy escrow託管賠償: Money set aside in advance — engagement contracts require vendors to pre-fund it — so that when the system causes harm, compensation and rollback are already funded rather than merely promised. Community evaluators are paid from it, and Pack 6 designs decision traces to settle against it as civic receipts returning value to knowledge custodians. A design commitment, not yet a running settlement; where financial stakes are low, pause triggers replace escrow.
perspective receipts視角回執: The Attentiveness pack’s acknowledgement artefact: a receipt that lets each contributor find and correct how their words were represented in the record. Distinct from Pack 6’s proposed financial civic receipt; the buildable instrument is specified on the Measures page.
data coalition資料聯盟: In Pack 6’s proposed settlement design, an existing community institution — a neighbourhood association, tribal council, union, craft cooperative, or religious congregation — acts as steward of community-held data and models its members authorise it to manage. It bargains the Engagement Contract, decides what may be shared and what stays offline, and would receive settlement collectively rather than per person. Prefunded escrow could back draws and settlement under the charter; valuation and settlement windows remain open design questions. Its mandate cannot be assumed from the institution’s existence.
shared eval registry共享評測登錄庫: A public, Wikipedia-like registry of versioned, community-authored evaluations: affected people write the tests for harm and repair, and passing them becomes a release gate. Weval.org is the working example; the instrument specification lives on the Measures page.
WevalWeval: The working shared eval registry: a public, Wikipedia-like registry of versioned, community-authored evaluations built by the Collective Intelligence Project, where affected people write the tests for harm and repair and passing them becomes a release gate. CIP’s case — from red lines to green lines — is that harm-avoidance benchmarks reward mediocre care. One of us (Audrey) is a senior research fellow there. See Sources.
Reinforcement Learning from Community Feedback (RLCF)社群回饋強化學習（RLCF）: A reinforcement-learning approach in which approved community evaluations — authored by affected people rather than lab raters alone — supply the feedback signal for model updates; the same evaluations can also steer routing or policy without retraining. RLCF is an implementation choice that follows from the Responsiveness pack’s more basic civic commitment: the community defines what counts as harm, repair, and improvement before any technical feedback loop begins.
shadow mode / Apprentice Model影子模式: The first stage of the Apprentice Model: the system sees real inputs and proposes real actions but does not act, its proposals compared against the human or prior system so divergences can be studied before trust is granted. Shadow mode, then canary release, then guarded full rollout — apprenticeship as legitimate peripheral participation (Lave and Wenger, 1991), read as the mechanism by which a room grants a system real work. Specified on the Measures page.
canary release金絲雀發布: The second stage of the Apprentice Model: deployment to a small, stratified slice of real cases with automatic rollback triggers if drift exceeds defined bounds. A canary is a probe, not a stand-in for the public — canary health and coverage are Pack 3 supporting diagnostics. Specified on the Measures page.

Diagnosis

Architecture

Design

Instruments

The Trace