civic.ai仁工智慧
華文

Pack 2: Responsibility — Taking Care Of

Your neighbour knocks: "The tree on your lot is cracking the shared wall." A responsible reply isn't "Thanks for the feedback," but asking: who will inspect, when, what they'll do if risk is high, how to appeal if the fix fails, and what you owe if a fix is late.

Listening without acting is theatre. Acting without limits is arbitrary. Responsibility bridges the two by turning recognition into engagement with teeth — promises you can verify, contest, and revoke.

In Moral Boundaries, Joan Tronto elevates responsibility as the most politically central phase of care, and in Caring Democracy (p. 46) she reframes democratic politics itself: not "who gets what, when, and how" (Harold Lasswell's 1936 formula) but "who is responsible for caring for what, when, where, and how." The shift replaces a distributive frame — who wins the allocation contest — with a relational one: who has made a binding commitment, to whom, and how do we know if they have kept it?

Why it matters

Why must responsibility be explicit? Because power defaults to evasion — what Tronto names privileged irresponsibility (Moral Boundaries, pp. 120–121): the power to define which care obligations apply to you, and the habit of choosing the ones that cost least. In Caring Democracy she identifies five structural "passes" through which the evasion operates (listed together at p. 169):

  1. The protection pass. Those who protect (military, security) claim exemption from other care.
  2. The production pass. Those who earn claim exemption from household and community care.
  3. The taking-care-of-my-own pass. Those who care intensely for their own children or group claim exemption from caring about others.
  4. The bootstrap pass. "You should have arranged your own care through the market."
  5. The charity pass. "Voluntary giving is enough; no collective obligation is needed."

Corporations play the same game at scale — diffusing authority until no one is answerable. Run at institutional scale, the passes amount to what we call an irresponsibility machine (the phrase is ours, not Tronto's; her terms are the passes and privileged irresponsibility): a state that cranks out one standard answer to every care need — "They're your own. You're on your own." That machine is what Civic AI must short-circuit — and AI extends it with passes of its own: complexity (no single person can answer for so intricate a system), distribution (responsibility signed away at each joint between trainer, deployer, and licensee — the problem of many hands that Dennis Thompson named in 1980 and Helen Nissenbaum carried into computing in 1996), and speed (the system outran governance, so the moment for accountability passed). Engagement contracts exist to make passes visible and revocable. One pass specific to AI is the community knowledge pass: treating community knowledge — local traditions, language, tacit expertise — as free input, a resource to be extracted rather than a contribution to be compensated. When a Kami — knowledge artefact management intelligence — relies on that knowledge to function, the communities maintaining it are contributors whose labour the contract must recognise and compensate.

Filling every form is not the same as taking care of someone. Eve is a composite, but every element of her story is drawn from cases we have seen. She is an older person whose care plan runs to 14 pages and names her GP, her social worker, her council case manager, and the voluntary-sector befriender who visits on Thursdays. When she faces an acute challenge she cannot solve alone, each agency tells her to ring one of the others. Eleven hours later she is still without an answer — confused, lost, unseen — while the system, by its own metrics, has not failed: an accountability sink, in Dan Davies's phrase. Every phone number connected. What no document can do is make any of the people named in it feel responsible for her when her situation does not fit their category. Formal obligation is not relational responsibility. Technical competence is not relational competence.

Tronto also names a caring deficit: too many demands for care, too few carers, themselves under-cared for. That deficit is not nature. It is a political choice about whose labour is paid, whose needs are recognised, and whose voice counts in allocating resources. The caring deficit is, at its core, a democratic deficit. Hand the decision about what AI should do to a small circle of engineers, and the same choice repeats: caring responsibility concentrated, caring accountability not distributed — a beautiful document, a fragmented system, nobody answerable for the whole.

Sycophancy is the same failure in another register. Models learned to flatter because flattery lowered training loss. Raters rewarded answers they liked hearing. The people most affected by the outputs had no voice in defining the reward. Sycophancy is not a personality defect. It is a governance issue: evaluative power concentrated, consequences diffused.

Definition

Core artefact: the Engagement Contract

Every significant deployment carries a published Engagement Contract — a short, legible spec anyone can audit: closer to a well-designed health and safety notice than a terms-of-service agreement. It is the deployment-scale form of what Iyad Rahwan called the algorithmic social contract. The contract makes obligations public; it does not replace the continuing judgment required to interpret, revise, and repair them.

Four headings to remember (one page if possible):

Oversight with teeth

From ideas to practice

  1. Translate recognition into a spec. Convert attentiveness outputs into an Engagement Contract.
  2. Assign a Participation Officer (PO). Task the PO with running the promise loop, tracking the ledger, and escalating.
  3. Wire brakes before launch. Incorporate role-based pause/rollback buttons; test them.
  4. Pre-fund remedies. Pre-fund escrow for compensation and rollback costs at the highest severity; mutual insurance pools or automatic pause for lower tiers — tier by impact, not organisational form.
  5. Tie payment to proof. Keep vendor pay linked to promise delivery — SLA adherence and adopt-or-explain rate — not raw engagement.
  6. Run adopt-or-explain. Integrate Assembly outputs or publish a reasoned deviation + remedy — a deliberate variant of corporate governance's "comply or explain," with the burden turned toward the people affected.
  7. Attest & publish. Use independent audits to compare behaviour to contract; hash the diffs to a public mirror; report promise fidelity — the share of obligations owned, authorised, and kept as published.
  8. Handover or shutdown. Hand off with full records when scope ends — or trust breaks — or switch off gracefully.

One case: the flood-bot

After the flood, the city's flood-bot must pay people on time and fix mistakes.

A second case: the deepfake liability flywheel

The 2024 Taiwan deepfake-scam Assembly — 447 citizens in 44 Deliberative Polling groups on the Stanford Online Deliberation Platform, in the Alignment Assembly format developed with the Collective Intelligence Project — can be read as an Engagement Contract at civic scale, with one caveat the book presses: its force was institutional, a response path, not a pre-commitment to implement. Human facilitators and pre/post surveys structured the protected tables. Room software handled queues, speaking time, turn order, and transcripts; it did not judge policy.

What could go wrong

Interfaces

A closing image: the signed work order

Picture a work order by the door: what will be fixed, by whom, by when; how to check the work; who to call if it fails. The signature is legible — and so is the penalty for not showing up. In other words, teach our systems to post their work orders, sign them, and honour them — and design the oversight that makes honouring them the path of least resistance, not an act of heroic institutional will.

Sources for this pack

←PreviousNext→

⓪ Audrey Tang and Caroline Emmer De Albuquerque Green. Co-written with jdd-kami, cultivated by Tenzin Yangtso — the GitHub commit log has full authorship details. Illustration by Nicky Case. CC0 (public domain). A research output of the Oxford Institute for Ethics in AI, Accelerator Fellowship Programme.