
Start in your room
Before choosing a system, ask, “Who keeps this, who can say no, and when do we review or end it?” Agree on its scope, name the people with authority to stop it and arrange a handover if the keeper can no longer serve. Work through the four readiness questions, including an end date. The community may choose no AI or no Kami, and any service duty must continue elsewhere when a component retires.¶ 1 (permalink to paragraph 1)
We borrow a spiritual image from Shinto practice: a kami linked to a place, such as a river or a grove (Thomas Kasulis’s Shinto: The Way Home is the account we lean on). In our metaphor, it tends that place and leaves the rest of the forest alone. Even its house is held lightly: at Ise the shrine is rebuilt every twenty years, and the kami is ritually carried into the new one. The building is renewed; the relationship to the place continues.¶ 2 (permalink to paragraph 2)
For an institutional Kami, the image becomes a design requirement. Its limits have to be engineered: resource caps, end-date timers, non-expansion pacts and succession duties.¶ 3 (permalink to paragraph 3)
Saying a Kami recedes when its work is done is too easy. Useful systems tend to become standing infrastructure, and vTaiwan, Polis, Monlam AI, ROOST and Engaged California all point toward continuity, not vanishing. Nobody should assume a good system retires on its own. Retirement has to be written into a contract: a named authority who can shut it down, an end or review date, non-expansion rules, resource and retention caps, handover tests, and a public record of succession.¶ 4 (permalink to paragraph 4)
But caps and timers only limit what a system does at runtime. They do not reshape the training and update processes that teach it what to preserve, resist or ignore. If those loops reward persistence, scope-seeking or resistance to correction, governance has to step in earlier: in how feedback is gathered, whose evaluations count, and what limits are set before each release and revision. So Pack 6 also covers process. Engagement contracts, community-authored evaluations and feedback loops keep a local Kami from hardening into a centre of power.¶ 5 (permalink to paragraph 5)
Symbiosis is the meta-level rule of the 6-Pack: even well-governed care can become dangerous if it hardens into permanent rule. The term is used in its broad biological sense, the living-together of unlike organisms, which can run anywhere from parasitic to mutualistic. Pack 6 is the gardening that holds the relationship at the mutualistic end and keeps it from sliding into dependency or permanent rule.¶ 6 (permalink to paragraph 6)
The word has a history in computing too. J. C. R. Licklider’s “Man-Computer Symbiosis” (1960) imagined human and machine thinking together, with the human setting the goals. Pack 6 reclaims his word for a civic register: symbiosis with a community, not with one operator. (Taniguchi and colleagues give this intuition a theoretical form in Symbiotic Alignment via Collective Predictive Coding: A Theoretical Framework for Co-Creative Human–AI Ecosystems, where agents with limited scope negotiate shared meaning through collective predictive coding rather than converging on a single master model.) Pack 6 keeps care local, limited, plural, and temporary.¶ 7 (permalink to paragraph 7)
One clarification bears repeating: the Kami does not care. It is infrastructure through which human communities care for each other. A hospital bed is not a carer, but it is part of a caring system. That is the point technological mediation theory makes about artefacts in general, and the verdict Aimee van Wynsberghe reached for a robot that replaces the care-giver: on Tronto’s terms it “does not meet the requirements of a good carer” (2013, p. 427). The moment a Kami is treated as the carer itself, rather than the scaffolding that makes human care possible, it has crossed from symbiosis into the displacement of caring labour.¶ 8 (permalink to paragraph 8)
Definition
We adapt Tronto’s fifth phase as solidarity and propose symbiosis as a sixth for keeping AI within human communities.¶ 9 (permalink to paragraph 9)
Symbiosis asks whether a system behaves like a local Kami (the word came first; the initials, knowledge artefact management intelligence, caught up) or like a permanent centre of power.¶ 10 (permalink to paragraph 10)
- Constitutional boundedness. Every AI system has a stated purpose, resource caps and a sunset date, all specified in advance.¶ 11 (permalink to paragraph 11)
- Service duty survives the component. The component retires; the public obligation hands over.¶ 12 (permalink to paragraph 12)
- Non-expansion pact. No AI system may widen scope without fresh authority and local consent.¶ 13 (permalink to paragraph 13)
- Treaties over hierarchies. Shared protocols let limited-scope systems cooperate without one system ruling them all.¶ 14 (permalink to paragraph 14)
- Subsidiarity with escalation. Local first, and escalated only when the local unit cannot fix the problem. The principle was named in Quadragesimo Anno in 1931 and later adopted by European Union law.¶ 15 (permalink to paragraph 15)
Our design expectation is that lifting a cap requires independent oversight approval and extending an end date requires fresh community authority. These are stronger constraints than a cap or date the operator can lift alone, but neither is a safety proof.¶ 16 (permalink to paragraph 16)
Why it matters
Packs 1–4 describe how care should be practised, and Pack 5 how it scales. Pack 6 answers a different question: what stops caring systems from centralising into a new permanent centre?¶ 17 (permalink to paragraph 17)
That is why symbiosis is not optional polish. A system can be attentive, responsible, competent, responsive, and solidaristic inside its lane while still becoming too entrenched to replace. Symbiosis keeps “useful” from becoming “indispensable.” The danger comes from success, and it needs no malice. Useful systems attract dependency, dependency creates leverage, and leverage accumulates. The test, at every stage, is whether the community keeps the practical ability to leave. The legal right is not enough; it needs the operational capacity too.¶ 18 (permalink to paragraph 18)
Technical limits alone do not keep care local. Institutional roots do. Existing intermediate institutions (churches, unions, neighbourhood associations, cultural traditions, local governments) are not stakeholders to be consulted. In Peter Berger and Richard John Neuhaus’s phrase, they are the mediating structures of public life, the primary actors in care. The Kami is scaffolding for their participation: infrastructure that lets a temple or a cooperative join decisions that affect it. It replaces neither. A deployment that bypasses the institutions closest to the community has violated subsidiarity before it begins.¶ 19 (permalink to paragraph 19)
These institutions act as custodians, not consultees: a union can veto changes to the triage system its members work alongside; a neighbourhood association reviews the River-Steward’s data retention; a cultural organisation mandates community-authored evaluations in its language.¶ 20 (permalink to paragraph 20)
Custodianship also has an economic expression. When a Kami draws on a community’s knowledge (an elder’s context, a translator’s nuance, a craft cooperative’s practice), the decision trace it already keeps doubles as a civic receipt, designed to settle against a pre-funded escrow that returns value to the community’s data coalition rather than extracting it. The coalition descends from the data-as-labour argument and the mediators of individual data proposed in Radical Markets. The trace half of that mechanism runs today; the settlement half does not yet run anywhere we know of. Value flows back to the soil that grew it.¶ 21 (permalink to paragraph 21)
Data as soil, not oil
The extractive economy treats data as oil: located, pumped, refined in centralised facilities, and depleted, with the value accruing to whoever owns the refinery. The metaphor fails at its root, as Nigel Shadbolt argued in 2016: data is not a rival good, and is better treated as shared infrastructure.¶ 22 (permalink to paragraph 22)
A care-based civic model, following Tenzin Yangtso’s phrase, treats data as soil: a living, shared substrate that must be locally nurtured, kept free of toxins that bioaccumulate, and never exhausted by over-farming. Over-farming does not exhaust the data, which use does not deplete. It exhausts what grows the data: the trust, consent and willingness to contribute of the people who keep it. The metaphor changes what we optimise for. The extractive question is how to get maximum value; ours is how to keep the ecosystem healthy. We also ask how to stay locally responsive and how to make sure communities can leave.¶ 23 (permalink to paragraph 23)
Soil implies local models. Quantised open-weight models now run on hardware a community can own: a laptop in a community centre, a small server in a council office, or a dedicated device in a clinic. The AI runs in the room; when the meeting ends, the model stays. Escalation to frontier capability follows subsidiarity: a specific request for a specific capability, logged with reason and duration, after which local control resumes. A care ethic that ignores the environmental cost of its infrastructure is not fully attentive.¶ 24 (permalink to paragraph 24)
What it looks like in practice
- Contract as code. The Engagement Contract travels with the deployment as a public, machine-readable rulebook encoding scope, consent rules, portability, and shutdown duties.¶ 25 (permalink to paragraph 25)
- Resource caps. Compute, reach, and retention are capped; exceeding caps triggers pause and review.¶ 26 (permalink to paragraph 26)
- Federation treaties. Peers agree on exchange formats, rate limits, safety pacts, and appeal handoffs across boundaries.¶ 27 (permalink to paragraph 27)
- Succession plans. Institutional records, evals, and aggregate traces transfer; private interaction histories do not. Sometimes the right ending is transfer of duty to a successor, not deletion.¶ 28 (permalink to paragraph 28)
From ideas to practice
- Write bounds as code. Put purpose, caps, and sunset conditions in the Engagement Contract so infrastructure can check them.¶ 29 (permalink to paragraph 29)
- Sign treaties. Join federations with machine-readable terms for sharing, dispute, repair, and appeal.¶ 30 (permalink to paragraph 30)
- Run exit drills. Practise handover twice a year and verify portability and continuity. Exit readiness is measured in the drill, not in the crisis.¶ 31 (permalink to paragraph 31)
- Escalate by subsidiarity. Escalate only when the local Kami cannot handle life-and-safety or livelihood harms; log why and for how long.¶ 32 (permalink to paragraph 32)
- Retire with honours. Archive traces, evals, and lessons so the next Kami starts stronger.¶ 33 (permalink to paragraph 33)
Buildable tools
- Limit enforcers for compute, reach, retention, and policy.¶ 34 (permalink to paragraph 34)
- Treaty registry for discovery and compliance checks.¶ 35 (permalink to paragraph 35)
- Local safeguard stack for community-run models and detectors: GPT-OSS and gpt-oss-safeguard for bring-your-own-policy moderation, and ROOST-style open tools for detecting child-safety harms, without centralising authority.¶ 36 (permalink to paragraph 36)
- Succession kit with handover scripts, fidelity checks, and cold-start playbooks.¶ 37 (permalink to paragraph 37)
- Ecology dashboard for diversity, redundancy, exit-ease, and handover readiness across Kamis.¶ 38 (permalink to paragraph 38)
One case: the flood-bot
This fictional composite is a teaching illustration, not an actual deployment or measured result. Its 90-day record-retention limit illustrates a contract choice, not a retention mandate or default.¶ 39 (permalink to paragraph 39)
- Boundedness. The River-Steward’s contract caps scope to post-flood relief for six weeks; the record-retention limit is 90 days unless individuals opt in to transfer.¶ 40 (permalink to paragraph 40)
- Treaty. The city joins a regional aid federation treaty on the River-Steward’s behalf, covering shared formats, safety alerts, and appeal handoff.¶ 41 (permalink to paragraph 41)
- Subsidiarity. When a cross-border housing issue arises, the bot escalates to the regional Kami; the handoff reason and duration are logged.¶ 42 (permalink to paragraph 42)
- Retirement. On week six, the River-Steward hands maps, tests, institutional records, and the model itself to the housing office exactly as the contract stipulated. The switch-off is logged, the ledger archived, and continuity is tested before shutdown.¶ 43 (permalink to paragraph 43)
What could go wrong
- Imperial creep. A capable system seeks new domains. Fix: Hard caps; non-expansion pact; fresh authority for scope changes.¶ 44 (permalink to paragraph 44)
- Within-scope power accumulation. A system can become dangerous without formally changing scope if it gains too much leverage inside its own lane. Fix: Caps verified by external infrastructure, not self-reported; independent resource audits; capability monitoring by the oversight board.¶ 45 (permalink to paragraph 45)
- Treaty fragmentation. Too many standards. Fix: Minimal core, adapters, conformance tests; polycentric but interoperable.¶ 46 (permalink to paragraph 46)
- Zombie agents. No one turns systems off. Fix: Sunset by default; alarms; “no attestation, no runtime.”¶ 47 (permalink to paragraph 47)
- Builder attachment. Builders treat limited-scope agents as extensions of their identity and resist retirement. This is Robert Michels’s iron law of oligarchy in miniature. Fix: Term limits for named keepers; rotate spokespeople; separate authority from personal brand; treat succession as a governance requirement, not a contingency plan.¶ 48 (permalink to paragraph 48)
- Receipt farming. Once civic receipts settle money, a Kami can be tuned to over-retrieve, or a coalition tempted to feed it low-value knowledge, until the escrow drains without care improving. Fix: Settlement at coalition level for a sum negotiated in the Engagement Contract; payout thresholds rather than unbounded per-use rates; receipts audited as competence records (Pack 3).¶ 49 (permalink to paragraph 49)
Interfaces
- Meta-level role. Pack 6 does not replace Packs 1–5; it keeps them from becoming their own negation: attentiveness without limits becomes surveillance; responsibility, a contractual trap; competence, indispensability; responsiveness, dependency masquerading as care; solidarity, infrastructure no one controls and everyone depends on. The rule guards itself as well: symbiosis without service continuity becomes abandonment in the name of purity.¶ 50 (permalink to paragraph 50)
- From Competence/Responsiveness (Packs 3, 4): only competent, responsive systems earn the right to keep serving.¶ 51 (permalink to paragraph 51)
- From Solidarity (Pack 5): treaties, IDs, and portability make symbiosis feasible. Care practices must embody democratic values all the way down, not merely produce good outcomes, so those treaties carry justice, equality, and freedom as operating constraints. A treaty that coordinates efficiently but excludes minority communities from its governance is capture wearing solidarity’s clothing.¶ 52 (permalink to paragraph 52)
- To Attentiveness (Pack 1): retired systems gift maps, evals, and operational records to the commons: better first looks next time.¶ 53 (permalink to paragraph 53)
- To Responsibility (Pack 2): limits and end dates are contractual.¶ 54 (permalink to paragraph 54)
A closing image: the river guardians
Imagine a river tended by local guardians; each keeps its bank, shares warnings upstream and down, and steps aside when the season changes. The river does not need one ruler. It needs many guardians who know their stretch and know when to let go. Knowing when to let go is a moral competence more than a technical one. That discipline binds the makers too. A guardianship that cannot survive its founders fading is not care. It is dependency.¶ 55 (permalink to paragraph 55)
Sources for this pack
- Licklider, “Man-Computer Symbiosis” (1960): the word, and the human at the wheel.¶ 56 (permalink to paragraph 56)
- Kasulis, Shinto: The Way Home (2004): the kami the name echoes.¶ 57 (permalink to paragraph 57)
- Verbeek, “Materializing Morality” (2006); van Wynsberghe, “Designing Robots for Care” (2013), p. 427: why infrastructure can be part of care without caring.¶ 58 (permalink to paragraph 58)
- Berger and Neuhaus, To Empower People (1977): mediating structures as the primary actors.¶ 59 (permalink to paragraph 59)
- Michels, Political Parties (1911): the iron law of oligarchy.¶ 60 (permalink to paragraph 60)
- Quadragesimo Anno (1931), §79; Føllesdal, “Subsidiarity” (1998): where subsidiarity comes from.¶ 61 (permalink to paragraph 61)
- Ostrom, Governing the Commons (1990), p. 90: the eight design principles the custodial model maps onto.¶ 62 (permalink to paragraph 62)
- Shadbolt, “Data as Critical National Infrastructure” (2016); Arrieta-Ibarra et al., “Should We Treat Data as Labor?” (2018); Posner and Weyl, Radical Markets (2018): data as infrastructure, as labour, and the data coalition.¶ 63 (permalink to paragraph 63)
- Jackson, “Rethinking Repair” (2014); Russell and Vinsel, “Hail the Maintainers” (2016); Puig de la Bellacasa, Matters of Care (2017): maintenance and repair as the substance of care for technical systems.¶ 64 (permalink to paragraph 64)
- Hui, The Question Concerning Technology in China (2016): technodiversity; Drexler, Reframing Superintelligence (2019): bounded services rather than a single agent.¶ 65 (permalink to paragraph 65)
- The full trail is on the Sources page.¶ 66 (permalink to paragraph 66)
