At a crosswalk, drivers slow for a child. No one stops to solve an equation. A need appears; a duty follows. That’s attentiveness.¶ 1 (permalink to paragraph 1)
Now scale up. An AI looks at a world full of “crosswalks”: workers, rivers, languages, communities recovering from floods. It can treat each as an obstacle or as a relationship asking for care. The difference begins with the first look.¶ 2 (permalink to paragraph 2)
In Caring Democracy (2013, p. 34), Joan Tronto calls attentiveness “a suspension of one’s self-interest, and a capacity genuinely to look from the perspective of the one in need.” The opposite is what she names privileged irresponsibility: the luxury of not noticing. Privilege, Tronto writes, grants “the opportunity simply to ignore certain forms of hardships that they do not face; I suggest that we call this form of privilege ‘privileged irresponsibility.’” Attentiveness is the discipline of refusing that luxury.¶ 3 (permalink to paragraph 3)
Carol Gilligan, in In a Human Voice (2023), names the muscle this trains: “Radical listening holds a potential for transformation because it starts from a place of not knowing and develops the muscle of curiosity.” Curiosity is built by turning, again and again, toward whoever is missing.¶ 4 (permalink to paragraph 4)
This vocabulary is inherited. Berenice Fisher and Joan Tronto first set out the four phases of the care cycle in 1990: caring about, taking care of, care-giving and care-receiving. Each of Packs 1–4 is named for the moral element Tronto assigned to its phase. Others have carried them into technology before: Aimee van Wynsberghe’s care-centred value-sensitive design (2013) used the same four phases and moral elements to design and evaluate care robots.¶ 5 (permalink to paragraph 5)
Packs 1–4 keep her question: how is each element carried out with and without the machine, and by whom? They widen the unit from one artefact in one ward to a deployed system answerable to the people it serves. The Sources page traces every such debt by pack.¶ 6 (permalink to paragraph 6)
Broad listening, bridging maps and perspective receipts help people notice relationships and gaps. The noticing itself still takes human judgment.¶ 6a (permalink to paragraph 6a)
Start in your room
Take the notes from your last meeting, or the summary an AI gave you, and ask: “Whose experience is missing from these notes?” Compare who is affected with who is represented. The difference is your representation gap. Then invite the people left out to correct the record in their own words, because an AI can sort what was said but cannot fill in a voice that never spoke.¶ 8 (permalink to paragraph 8)
Definition
- Relationships first. The relationship is the basic unit of care. Roles and dependencies make some duties visible. Listening means secure channels where people give input voluntarily and are recognised. Surveillance is attention without relationship; listening is attention within one.¶ 9 (permalink to paragraph 9)
- Power must answer questions. Decisions should be explainable and challengeable. If no one can question you, the process isn’t fair.¶ 10 (permalink to paragraph 10)
- Be precise only when it helps. Start with stories and people. Add numbers when they clarify; update them when reality changes.¶ 11 (permalink to paragraph 11)
- Rights baseline. This project operationalises care ethics within rights-respecting, pluralist democratic institutions. We use the Universal Declaration of Human Rights plus local constitutional rights as the floor for participation. Rights set the threshold for standing in a relationship, and the floor that keeps care from becoming domination.¶ 12 (permalink to paragraph 12)
- No fake pluralism. If someone’s basic existence is being erased, they cannot participate as an equal in a bridging process. Claims that try to erase someone’s standing are recorded but do not set the agenda.¶ 13 (permalink to paragraph 13)
Why it matters
Many AI plans try to “learn the objective” from old data. But shared goals are bargains among changing lives, and when people who were ignored finally speak, the target moves. A fixed goal guessed in advance will fail. A deliberation system also needs to know when a voice is human, affected and accountable enough to count, and AI makes that harder to know when participation can be faked at scale.¶ 14 (permalink to paragraph 14)
Attentiveness offers another route: align AI to a trusted process that listens, explains, adapts and can be corrected. Summaries show their sources. Unknowns are stated. Anyone can ask for a revision, including new voices that arrive later.¶ 15 (permalink to paragraph 15)
Rule of thumb: Bridge first, decide second. For acute harms (life-safety, livelihood), default to reversible protection immediately while bridging proceeds in parallel.¶ 16 (permalink to paragraph 16)
What it looks like in practice
Look for absences. “We heard nothing from night-shift carers, so go find them.” Missing voices are evidence. Hélène Landemore, in a Yale News interview about her book Politics Without Politicians: The Case for Citizen Rule (2026), borrows G. K. Chesterton’s “jolly hostess” test of democracy: “all real democracy is an attempt (like that of a jolly hostess) to bring the shy people out”. The point is to bring the quiet people into the room.
In Democratic Reason: Politics, Collective Intelligence, and the Rule of the Many (2013) she argues that, under the right conditions, cognitively diverse collectives can outperform decisions made by few. The argument is conditional. It concerns diversity of perspectives, and a mini-public does not automatically represent the whole public or decide better every time. Open Democracy: Reinventing Popular Rule for the Twenty-First Century (2020) develops that into a case for open mini-publics. Random or stratified selection is a way to get cognitive diversity.
¶ 17 (permalink to paragraph 17)Show your work as a brochure. Present AI outputs as brochures: shared, editable artefacts that link every claim to its sources and mark disagreements. A private chat reply smooths them into false consensus.
¶ 18 (permalink to paragraph 18)Share attention fairly. Give time to those most affected, and extra attention to smaller groups that are easy to miss and hard to reach.
¶ 19 (permalink to paragraph 19)Read the conflict pattern. When mapping disagreement, ask whether the same groups oppose each other on every issue (reinforcing cleavages) or whether alliances shift from issue to issue (cross-cutting cleavages, the bridges worth finding). The terms come from Lipset and Rokkan. The map should show who can bridge the divides and flag the splits that need extra care.
¶ 20 (permalink to paragraph 20)Build in repair. Sunset clauses, review points, reversible defaults, and the humility to shut down or hand off.
¶ 21 (permalink to paragraph 21)
From ideas to practice
- Listen widely. Take input by voice, text, and simple forms. Keep original language next to translations. Offer offline and accessible options.¶ 22 (permalink to paragraph 22)
- Map relationships and disagreements. Make a bridging map that shows where people agree, where they clash, and why, without forcing a fake average.¶ 23 (permalink to paragraph 23)
- Send receipts. Tell contributors where their words appear, and let them correct mistakes.¶ 24 (permalink to paragraph 24)
- Set a fair queue. Spend more time where harm is high and voices are quiet. Make the rules public.¶ 25 (permalink to paragraph 25)
- Decide with brakes. Require the map, the receipts, and an oversight check before big changes ship.¶ 26 (permalink to paragraph 26)
- Publish the gap. Report the representation gap before and after each cycle: which materially affected groups are still missing or badly under-represented in the record.¶ 27 (permalink to paragraph 27)
Buildable tools
- Broad listening. Multi-language, multi-channel input with source and uncertainty kept intact.¶ 28 (permalink to paragraph 28)
- Bridging maps. Charts of overlap and disagreement, with citations, using tools like Polis (Small et al., 2021) to surface consensus without flattening difference. Aviv Ovadya named the design principle bridging-based ranking.¶ 29 (permalink to paragraph 29)
- Perspective receipts. Receipts that allow each person to find and correct how they were represented.¶ 30 (permalink to paragraph 30)
- Machine-checkable rules. Community data rules written so software can enforce them automatically. This is one use of the policy-as-code guardrail engine Pack 3 builds.¶ 31 (permalink to paragraph 31)
- Fair queues. Simple algorithms that favour high-risk issues and quiet voices.¶ 32 (permalink to paragraph 32)
- Jolly-hostess prompts. Governing instructions that tell the AI to act as a bridge: represent quieter groups, flag under-representation and show where crossings may be possible. Attentiveness then becomes part of the structure.¶ 33 (permalink to paragraph 33)
One case: the flood-bot
This is a fictional composite, an illustration rather than operating evidence.¶ 34 (permalink to paragraph 34)
A midsize city is hit by floods. The city launches a simple chatbot, the flood-bot, to help people apply for emergency cash. It decides routine claims under the city’s published eligibility rules and routes the cases those rules flag to a caseworker.¶ 35 (permalink to paragraph 35)
The same composite runs through all six packs, inside the six weeks its contract allows: listening here, a contract and a pause in Pack 2, a replacement rule trialled in Pack 3, a repair in Pack 4, wider defences in Pack 5, and retirement in Pack 6.¶ 35a (permalink to paragraph 35a)
- Listening. People send voice notes, texts, or visit a kiosk. Messages stay in the original language, with a clear translation beside them. Elena is a night-shift hospital worker. Her landlord never gave her a formal lease, and her papers were lost to the flood. At a kiosk she records a voice note in her first language: she has no standard documents, but neighbours and a landlord can vouch for her.¶ 36 (permalink to paragraph 36)
- Mapping. The team sorts needs into categories: housing, wage loss, documentation barriers. They keep disagreements visible: renters and homeowners need different proofs.¶ 37 (permalink to paragraph 37)
- Receipts. Every contributor gets a link to see how their words were used and a button to say “that’s not what I meant.” A named person answers each correction within a published response window.¶ 38 (permalink to paragraph 38)
- Fair queue. People flagged as both high-need and low-digital-footprint, the ones standard channels most often miss, get priority human review. The queue decides nothing about the case. It makes sure a person looks at it promptly.¶ 39 (permalink to paragraph 39)
- Measure. A public dashboard shows who has been reached, who is still missing, and whether the queue is fair.¶ 40 (permalink to paragraph 40)
What could go wrong
One metric runs the show. Engagement is up, but trust is down. Fix: Use a small set of balanced measures and rotate which one leads.
¶ 41 (permalink to paragraph 41)Listening theatre. Glossy reports, same outcomes: the tokenism rung of Arnstein’s ladder. Fix: Real decision gates, outside veto power, and spot checks.
¶ 43 (permalink to paragraph 43)Loud voices take over. Well-funded groups flood the channel. Fix: Throttles, fair quotas, public attention dashboards.
¶ 45 (permalink to paragraph 45)First movers freeze the frame. Early language locks in. Fix: Rolling windows and boosts for late but important views.
¶ 47 (permalink to paragraph 47)False equivalence. Treating harmful claims as equal. Fix: Separate facts from values, uphold basic rights, refuse fake equivalence.
¶ 49 (permalink to paragraph 49)Synthetic publics at scale. The hardest failure mode is fake participation at scale. In California, South Coast AQMD’s gas-appliance rulemaking drew more than 20,000 opposition comments sent through an advocacy platform that advertises AI tools. Reporting by the Los Angeles Times and by Patrick Sisson in Planning describes the episode as a warning that platform-generated comments can dilute real community voice.
The source describes a small sample: three of the five senders who replied said they had not written the comments, and the board rejected the proposal 7–5. That does not show the letters were AI-written or that the flood decided the vote. California’s SB 1159, signed into law in September 2026, targets the same gap: it excludes AI systems as members of the public while allowing natural persons to use assistive tools at ordinary human volume. Fix: Provenance, rate limits, accountable standing, selective-disclosure identity (Pack 5), and audit trails that separate expression from amplification. Speech is a human right. Algorithmic amplification is not.
¶ 51 (permalink to paragraph 51)Capture by power. Oversight migrates toward those with most to lose from scrutiny. Fix: Keep oversight independent, rulings public, funding transparent, membership regularly renewed.
¶ 53 (permalink to paragraph 53)Procedural capture. Process professionals learn the format; resourced orgs dominate the queue. Fix: Fund participation — paid time, childcare, translation, community intermediaries — and track who is missing as well as who shows up.
¶ 55 (permalink to paragraph 55)
Interfaces
- From Responsiveness (Pack 4): repair loops and newly discovered needs restart the cycle.¶ 57 (permalink to paragraph 57)
- From Symbiosis (Pack 6): retired systems hand on maps, evals and receipts, a better baseline for the next first look.¶ 58 (permalink to paragraph 58)
- To Responsibility (Pack 2): attentiveness hands over the who, what and why, along with flags on rights and unknowns. Without attentiveness, responsibility works in the dark. Without responsibility, attentiveness only watches.¶ 59 (permalink to paragraph 59)
- To Competence (Pack 3): high-caution areas become small, safe-to-fail trials.¶ 60 (permalink to paragraph 60)
- To Solidarity (Pack 5): fair attention and open challenge build cross-group trust.¶ 61 (permalink to paragraph 61)
- To Symbiosis (Pack 6): The system serves a place and time, and treats shutdown as success.¶ 62 (permalink to paragraph 62)
A closing image: the jolly hostess who can still say no
Picture a jolly hostess who welcomes each guest by name and makes space for their baggage. She also moves through the room to find the person standing alone by the wall and asks the question only they can answer. That’s attentiveness. And because some guests try to erase others, the hostess keeps a firm rule: hospitality within a rights-respecting home. Teach our systems to be jolly hosts: attentive, in no hurry to optimise, curious about who is not yet in the conversation. We will keep more of what’s precious and create more that’s shareable. A need appears; a duty follows.¶ 63 (permalink to paragraph 63)
Sources for this pack
- Fisher and Tronto, “Toward a Feminist Theory of Caring” (1990); Tronto, Moral Boundaries (1993), pp. 105–108 and 120–121 — the phase, the element, and privileged irresponsibility.¶ 64 (permalink to paragraph 64)
- van Wynsberghe, “Designing Robots for Care” (2013) — a documented technology-design precedent for carrying the four phases into machines.¶ 65 (permalink to paragraph 65)
- Chesterton, “The Travellers in State” (1909), via Landemore’s Politics Without Politicians (2026), discussed in a Yale News interview — the jolly hostess; Landemore, Democratic Reason (2013) and Open Democracy (2020) — the conditional cognitive-diversity argument.¶ 66 (permalink to paragraph 66)
- Ovadya, Bridging-Based Ranking (2022); Small et al., “Polis” (2021) — bridging maps.¶ 67 (permalink to paragraph 67)
- Arnstein, “A Ladder of Citizen Participation” (1969) — listening theatre as tokenism.¶ 68 (permalink to paragraph 68)
- Tessler et al., “AI Can Help Humans Find Common Ground” (Science, 2024) — evidence that a model mediator can find statements divided groups endorse.¶ 69 (permalink to paragraph 69)
- Schroeder et al., “How Malicious AI Swarms Can Threaten Democracy” (Science, 2026) — synthetic publics.¶ 70 (permalink to paragraph 70)
- The full trail, with what each source contributed, is on the Sources page.¶ 71 (permalink to paragraph 71)
