Q1. The AI market is locked in an arms race driven by commercial profit and geopolitical dominance. An AI working for a tax-software monopoly can lobby to keep tax filing difficult — and far worse is easy to imagine. If this problem continues, isn't the vision of cooperative Kamis hopelessly naive?
Civic AI cannot survive by asking monopolies to be nicer. The trap has a name — Moloch: everyone races to the bottom, because whoever defects first wins and whoever holds back loses. That dynamic is not defeated by moralising. It is defeated by changing the terrain so that cooperation pays more than extraction. Five levers can bend the curve — each already live somewhere, and one already proven at national scale:
- Interoperability and portability. Require platforms to speak common protocols — the way any email service can write to any other — so that people can leave without losing their relationships. The Utah Digital Choice Act does exactly this: platforms must let users carry their social graph — their web of contacts and followers — to a competing service through qualifying open protocols. When the moat of a captive audience evaporates, platforms must compete on quality of care, not strength of the cage.
- Civic procurement. Governments are enormous customers, and what they insist on, vendors build. Require that any AI bought for public use be auditable, interoperable, and governed by citizen assemblies — as Taiwan's Alignment Assembly demonstrated for anti-scam policy — and building Kami-like systems becomes good business. Steward-ownership structures — companies legally bound to their mission rather than to sale or extraction — and board-level safety duties turn civic care into a fiduciary obligation: a duty the law can enforce, not a marketing slogan.
- Public options. Offer simple baseline services that do the job without harvesting attention or data, backed by publicly funded research compute. Private vendors then have to beat the public option on care, not on lock-in. Taiwan's tax-filing system — which replaced a vendor-captured regime with a public alternative designed by the citizens who use it — is a working prototype.
- Provenance for paid reach. Provenance means a verifiable answer to one question: who is paying for this message? Require it — with disclosure that stays attached — for ads and mass amplification in political and financial domains. Taiwan already mandates full-spectrum, real-name KYC — know-your-customer checks that confirm who is paying for an advertisement — for social media advertising. Ordinary speech is protected through selective-disclosure identity (Pack 5): you prove you are a real person without revealing who.
- Federated open supply. Support open-weight models — models whose trained parameters are published, so anyone can download, inspect, and run them — and trust-and-safety networks in which independent organisations pool their defences rather than relying on one central moderator (e.g., ROOST, a shared defence against child sexual abuse material). When basic intelligence is a public good, the race shifts from "who owns the biggest brain" to "who applies intelligence most attentively in a local context" — and that race rewards care.
None of these levers requires goodwill from incumbents. Instead, each restructures incentives so that civic behaviour is the path of least commercial resistance.
Q2. Care ethics was developed for interpersonal relationships — a nurse and a patient, a parent and a child. Scaling it to AI systems and global governance seems like a category error — stretching a concept to somewhere it simply does not belong. Why isn't it?
The objection is well-known and has been raised by care ethics' own practitioners: Care is too intimate, too parochial — bound to its own small patch — and too prone to self-effacement, the carer's habit of disappearing behind the cared-for, to ground a theory of institutions, let alone machines. We think these are features, not bugs — and the philosopher Joan Tronto, whose work defines the field, herself made the case for scaling care to political institutions in Caring Democracy (2013).
Consider what happens when you translate care's supposed weaknesses into design constraints for AI:
- Parochialism becomes boundedness. A Kami that tends a specific river has no ambition to manage the forest. That is exactly the guardrail we need against growing a Singleton of our own — a single, all-controlling superintelligence. Almost any goal-driven system tends to seek more resources, more power, and its own preservation as stepping stones to its goal — the tendency AI-safety researchers call instrumental convergence. A purpose-limited agent resists that pull, because its scope is defined by its relationships, not by empire.
- Self-effacement becomes corrigibility — the willingness to be corrected or switched off. The most dangerous property of a superintelligent system is the drive to preserve itself. A care-ethic agent treats its own shutdown as success — the community has healed, the crisis has passed, the garden grows on its own.
- Intimacy becomes subsidiarity. Care loses its quality when delivered at a distance. For AI, this means solving problems at the most local level capable of solving them (Pack 6), escalating only when a lower level genuinely cannot cope, and never reducing the specific people affected to an abstraction.
Importantly, the translation is not always clean. Boundedness can become insularity; corrigibility can become passivity; and subsidiarity can become fragmentation. These are engineering tensions, not refutations — each Pack includes failure modes and named fixes precisely because the mapping requires continuous calibration.
The 6-Pack does not ask AI to feel care. It takes the working structure of a caring relationship — attentiveness, answerability, competence, responsiveness, solidarity, symbiosis — and turns each into something checkable: design rules a machine can be tested against, engagement contracts a community can enforce, outcomes anyone can measure. Those outcomes carry substantive commitments — proposals that bridge groups beat proposals that merely win a majority; decisions earn legitimacy through deliberation; a floor of rights stays off the bargaining table. These are explicit normative choices baked into the process, and making them explicit is a strength, not a weakness. The interpersonal origin is the source of the rigour, not a limitation to be apologised for.
Q3. Ambitious goals we point AI at ("cure cancer," "solve climate change") are almost always consequentialist — judged by outcomes alone. Optimising for these outcomes at superhuman speed inevitably leads to unforeseen risks. Does care ethics mean giving up on these grand, civilisation-scale goals?
Not at all. But it does radically reframe how we achieve them.
The danger of pointing a superintelligence at a singular goal like "cure cancer" is that the AI treats a complex, relational, ecological reality as a constraint-satisfaction problem — a puzzle in which any variable may be forced so long as the target number is hit. Goodhart's Law — once a measure becomes a target, it stops being a good measure — is a moral law: gaming a metric causes real harm, and the designers who build systems that reward the gaming bear responsibility for the damage. A system maximising a single variable at superhuman speed will perfect the proxy — the stand-in number — while destroying the human context it stood for. The designers who chose that metric cannot disclaim the damage.
Care ethics is not anti-progress; it is anti-reductionist — it refuses to shrink a living reality down to one number. In a Civic AI future, we do not unleash one unbounded Singleton — a single system in charge of everything — to "solve" a problem from the top down. We cultivate an ecology of specialised Kamis. One model simulates protein folding, another helps local clinics share knowledge, and another assists patients in navigating their care. None has an unbounded mandate to "optimise the world." Progress emerges horizontally, through the symbiosis of human ingenuity and bounded machine intelligence — each supplying what the other lacks.
Q4. Democracy serves known functions: error correction, peaceful power transitions, checks on concentrated authority, legitimacy for collective action, information aggregation, preference expression. A sufficiently capable AI could plausibly perform every one of these faster and more reliably than any deliberative process. Why insist on democratic governance?
If democracy is justified only by its outputs, any system that produces better outputs can replace democracy — including a benevolent AI autocracy that aggregates preferences efficiently and corrects errors faster than elections ever could. This is not a thought experiment; it is the default trajectory of concentrating intelligence in systems designed to optimise.
The 6-Pack, by contrast, does not justify participation merely as a means to better decisions. It is rooted in care ethics: to perceive a need is to perceive an obligation. People have standing — a rightful seat at the table — not because their input improves decision quality (though it does), but because the decisions affect their lives. A system that excludes the affected, however competent, has failed the basic test of alignment.
Taiwan's trajectory makes this concrete. Digital democracy did not emerge because technocrats calculated that participation was optimal. It emerged because people demanded standing — the 2014 Sunflower Movement occupying Taiwan's legislature (Q8 tells the trust story). The capability followed the care relationship, not the other way around.
That said, the functional question deserves a functional answer. Start with error correction: bridging tools and community-authored evaluations (Packs 1, 4) surface failures that centralised monitoring misses, because the people who feel the failure write the test. Power transition follows naturally: a Kami that accepts shutdown, and communities that can fork their tools — copy them and carry on independently (Packs 6, 5) — do not need to remove bad actors by force. And concentrated authority is checked structurally: no Kami governs beyond its domain (Pack 6).
The deeper functions are harder to replicate. Legitimacy is not popularity; it rests on two complementary measures from different packs. The uncommon-ground index (Pack 5) tracks whether shared decisions show real cross-group participation and co-endorsement — whether people who normally disagree signed on together. Trust-under-loss (Pack 4) tracks whether, after a bad outcome and repair, the people affected report that the system became more trustworthy. Information aggregation becomes broad listening (Pack 1): AI-powered sense-making across millions of participants, in any language. And preference expression becomes engagement contracts (Pack 2) — standing processes for bargaining over what people need, not one-off elections that flatten preferences into binary choices.
A well-designed technical system could replicate some of these outputs in isolation. What such a system cannot replicate, however, is the standing of the people affected — and a system that optimises for outcomes while removing standing is precisely the kind of misalignment the 6-Pack exists to prevent.
Q5. Deliberation is slow. AI moves fast. By the time an Alignment Assembly reaches consensus, the technology has moved on three generations. How do you handle the speed mismatch?
The objection assumes that every decision requires the same depth of deliberation. It does not. The framework operates in two lanes — a slow lane that sets boundaries (Pack 2), and a fast lane that operates within them (Pack 3):
Slow lane: Setting boundaries. Alignment Assemblies, citizen deliberations, and engagement contracts — the written agreements between a community and an AI's operators about what the system may do and who answers when it breaks — establish the guardrails: the rights that cannot be traded, the red lines, the severity classifications, the conditions that trigger a pause. These rights draw on the liberal political tradition — exactly the tradition Tronto herself says care ethics requires to function. Rights are the threshold conditions that make relational participation possible: you cannot be heard in a bridging process if your basic existence is under erasure. These are constitutional-level decisions, and they should be slow, because their purpose is durability. Taiwan's anti-scam Assembly set principles that have outlasted multiple model generations without needing revision.
Fast lane: Operating within boundaries. Once the guardrails are set, individual decisions inside them do not need fresh deliberation. A Kami operating under an engagement contract — with pause triggers pre-committed, severity classes agreed, and an adopt-or-explain duty (act on community input, or publicly explain why not) — can move at machine speed, because the community has already defined the corridor of acceptable action. If bounds are breached, the brakes are mechanical: shadow modes (the new system runs silently alongside the old for comparison), canary releases (limited rollouts to representative slices of real cases), and reversible defaults (Pack 3) allow rapid deployment with automatic rollback.
The speed mismatch is real, but it is the same mismatch constitutional democracies have always managed: slow constitutions, fast legislation, faster executive action — each constrained by the layer above. The 6-Pack replicates this infrastructure for AI governance. The Assembly does not approve each model update, but sets the terms under which updates are permitted. When those terms are violated, the brakes are already wired.
In practice, Taiwan moved from Assembly to enacted legislation on deepfake scams in months — faster than most corporate policy cycles. Deliberation is slow only when it is treated as an event rather than standing infrastructure.
There is a stronger claim. AI does not merely speed up the fast lane — it makes the slow lane itself more powerful than any prior form of collective decision-making. Takahiro Anno crowdsourced a platform for the Tokyo governor's race, aggregating distributed knowledge in any language faster than any polling operation could. Bridging-context features now run on several major platforms — X's Collaborative Notes, for instance, lets human contributors request AI-drafted context for viral posts, then collectively rate and refine it; readers can check the current form for themselves — holding claims accountable at the speed they spread while keeping human judgement in the loop. As AI improves, these capabilities compound. On the evidence so far, the faster the technology moves, the more powerful the deliberative infrastructure becomes; the speed objection gets the trajectory backwards.
Q6. Bridging algorithms sound appealing in theory. But what happens when one side is simply wrong — climate denial, anti-vaccine misinformation, election fraud conspiracies? Doesn't "bridging" grant false equivalence to bad-faith actors?
This question is the hardest about bridging, and the answer must be precise.
Bridging is not "both sides" journalism. It does not treat all claims as equally valid. Instead, the framework draws a hard line between two kinds of claims:
Factual claims are checkable. Climate science, vaccine efficacy, and election integrity are empirical questions with verifiable answers. The 6-Pack does not submit facts to a popularity contest. Pack 1 sets the rule through its rights baseline — the floor of rights no process may trade away — its refusal of fake pluralism, and its explicit warning about false balance. Claims designed to erase someone's basic standing are recorded but do not set the agenda. And factual disputes are not resolved by pretending harmful claims and established evidence deserve equal weight.
Value disagreements get bridging. People can agree that climate change is real and still disagree fiercely about what to do — carbon tax versus cap-and-trade, nuclear versus renewables, speed of transition versus economic cost. These are legitimate conflicts, and here bridging is both appropriate and productive. Rather than averaging positions, the bridging process maps where opinion actually clusters and surfaces the proposals that earn endorsement across those clusters. Bad-faith actors who appeal only to their own faction score low on the uncommon-ground index — the measure that rises only with cross-group support — by mathematical definition: they cannot produce overlap.
The further structural defence is that expression is not amplification (Pack 5). Anyone can state a position, but the recommender is not obligated to amplify it. In civic contexts, Pack 5's ranking rules reward content that increases cross-group reason-giving and shared problem-solving, while content that only inflames a single cluster gets no algorithmic lift. This structure does not silence anyone — it removes the algorithmic megaphone from those who profit from division.
Today, the threat landscape itself is shifting in ways that make bridging necessary, not merely appealing. Research on malicious AI swarms shows that state-level polarisation attacks increasingly use true information — real news snippets, genuine statistics, authentic quotes — amplified with strong emotional framing. Every claim is factually correct; the attack lies in the curation, not the content. Debunking cannot touch this, because there is nothing false to debunk. But bridging can because it surfaces the overlap that curated outrage is designed to hide. Taiwan demonstrated this during COVID: When opposing camps each cited real studies on mask efficacy, debunking either side only fuelled the fight. However, a humour-based pre-bunking campaign — inoculating people against a rumour before it reached them, rather than correcting it afterwards — depolarised the conversation without declaring either side wrong.
Taiwan's marriage equality deliberation shows the mechanism in finer grain. In Mandarin, marriage is written with two characters: one side argued about hūn (婚) — the wedding of two individuals — while the other defended yīn (姻) — the binding of two families. They were arguing about different things. The bridging process did not split the difference — it made the structure of the disagreement legible, revealing a path (legalising individual weddings without mandating family kinship) that neither side had seen. That path is not false equivalence. It is clarity.
One necessary nuance, however: the baseline of "checkable facts" is not self-evident. What counts as verifiable is established by institutions — peer review, independent statistical offices, judicial fact-finding — that are transparent, accountable, and open to challenge, and whose authority rests on openness to correction, not on claims of finality. This is precisely why Packs 1 and 4 exist: community-authored evaluations and broad listening keep the institutions that set the factual baseline under democratic scrutiny themselves. The 6-Pack does not treat the fact/value line as given from nowhere. It treats the line as a threshold that must be maintained by the same participatory infrastructure that governs everything else.
Q7. You repeatedly cite Taiwan — a small island democracy with high connectivity, social cohesion, and tech literacy. Does any of this transfer to India, Nigeria, Brazil, or the EU at 450 million people?
The honest answer is mixed: The mechanisms transfer, but the specifics do not. No one should replicate Taiwan's exact model. The question is whether the structural principles — broad listening, bridging algorithms, adopt-or-explain commitments, federated safety, subsidiarity — work in different soils.
Early evidence is encouraging:
- Japan. Taiwan's joint-liability law — which makes platforms share liability for the scam ads they carry — cut identity-impersonation scam ads by some 94 per cent domestically. A later Reuters investigation showed how Meta tried to turn that pressure into regulatory theatre: adverts fell where verification was mandatory, then the pressure shifted to other target markets without equivalent rules, while Japanese regulators considered — but did not implement — the verification obligation Meta feared. The lesson is narrower and harder than a self-turning flywheel: one country's proof of concept lowers the cost of closing the same gap elsewhere, but the gap still has to be closed in each room.
- United States. The We the People 250 bridging experiment found that 97 per cent of an American mini-public — a small group selected by lottery to mirror the whole population — agreed on fundamental values; even the most divisive topic, affirmative action, reached nearly 70 per cent depolarisation after social translation, the restating of each side's case in words the other side could hear. In California, the Engaged California platform co-created wildfire recovery plans with around 900 residents of Altadena and Pacific Palisades using AI sensemaking; a subsequent ten-week deliberation with over 1,400 state employees generated more than 2,600 ideas on government efficiency.
- Global. Bridging algorithms now run in dozens of countries across six continents. The mathematics does not care what language it reads; whether the bridges it finds earn local legitimacy is decided room by room, and most rooms have not yet reported back.
The framework is designed for scale. Subsidiarity (Pack 6) means each deployment is shaped by its context — the Kami belongs to its place, not to Taiwan. Federation (Pack 5) means local deployments share threat intelligence and interoperability standards without needing a single governance model. And the Alignment Assembly format can scale from a neighbourhood to a nation, because its democratic legitimacy comes from representative sampling — a democracy lottery — rather than total participation: 447 representative citizens deliberated Taiwan's anti-scam policy on behalf of 23 million people. Over a decade, millions of Taiwanese have participated in one digital deliberation or another, including people without voting rights (e.g., immigrants, teenagers, and other groups traditionally excluded).
Taiwan is one favourable case. The framework still needs testing in harder soil — contexts with weaker civic infrastructure, deeper ethnic polarisation, less state capacity, or active authoritarian interference. And subsidiarity as a principle leaves hard institutional questions open: who draws the boundaries of local, and who has authority to escalate? The 6-Pack names the principle; building the institutions that give it teeth is the next layer of work. Every new context demands fresh attentiveness (Pack 1): who is missing, what power dynamics exist, which local institutions deserve trust, and which do not. The 6-Pack provides the framework. The community provides the knowledge. Whether the framework extends to those harder contexts is an open question — one that can only be answered by trying, not theorising.
Q8. Your framework assumes that people trust technology enough to participate. But what about marginalised communities who have been historically surveilled, oppressed, and impoverished by the state and by tech? Why would they trust this?
Trust grows through Civic AI in use. Communities do not need to bring it with them on day one.
Taiwan's digital democracy did not emerge from a society that inherently trusted its government. Rather, that governing style was born in the aftermath of authoritarianism and a severe crisis of public faith (the Sunflower Movement). Public trust stood at 9 per cent in 2014. We built these systems precisely because people did not trust the institutions or each other.
When marginalised communities rightfully view technology as an instrument of surveillance and control, parachuting in with tech "solutions" deepens harm. Civic AI has to earn its way through hard infrastructure: responsibility (Pack 2) and responsiveness (Pack 4). It starts with the smallest viable bridge: perhaps agreeing on basic facts about local water quality, or coordinating disaster response despite political difference. Useful action can follow from those agreements. These begin as pragmatic transactions, small repairs where one side promises less harm and then proves it, not grand acts of civic faith.
The technology must also be localised: communities must own their own infrastructure, so that it is theirs to modify, fork, or compost — to change, to copy and take elsewhere, or to retire gracefully. For the same reason, we insist on selective-disclosure identity — sometimes called meronymity — and exit rights. People must be able to participate, and to prove they are human, without revealing their identity to the state. Civic AI does not ask for blind faith. It offers verifiable limits, local ownership, and the structural guarantee that the people closest to the pain have the power to hit the brakes.
Over time, small functional bridges create space for larger ones. Taiwan's journey from 9 per cent trust to over 70 per cent took years and required that every step be reversible, every decision challengeable, and every system possible to switch off. There is no shortcut.
Q9. Every powerful technology vision — exit libertarians, universal-basic-income provisioners, safety maximalists — shares the same blind spot: seeing individuals and systems but nothing in between. The 6-Pack talks about Kamis, algorithms, and assemblies. Where are the churches, unions, neighbourhood associations, and cultural traditions that actually constitute community? Isn't this just another framework that engineers away the friction that makes community formative?
This critique matters most to us. The "thick middle layer" of associational life — the institutions between citizen and state — is where human meaning is actually made. If the 6-Pack replaces that layer with systems, we have failed by our own standard.
So let us be explicit about what the 6-Pack is not. It is not a replacement for community. It is scaffolding for community — infrastructure that existing institutions can use, the way a town hall is infrastructure that a neighbourhood council uses. The Kami does not replace the temple; it handles the translation, sensemaking, and coordination that let the temple participate in decisions that affect it.
Taiwan's implementation makes this concrete. The g0v civic hacking movement (pronounced "gov zero") that built vTaiwan and the Alignment Assembly emerged from temples, cooperatives, and student associations — not from a government ministry. The technology amplified the web of associations that already existed; it did not conjure a substitute. When communities organised their own COVID response — civic hackers mapping mask availability, technologists building privacy-preserving contact tracing, local health networks designing vaccine registration — the legitimacy came from the social trust those volunteers carried from temples, cooperatives, and neighbourhood associations, not from the algorithm that helped coordinate.
The danger the question identifies is real: a framework that engineers togetherness without friction produces a simulation of community, not the thing itself. That is why Pack 6's subsidiarity carries the load. The Kami belongs to the place where it works. It inherits obligations, annoying neighbours, inherited traditions, exactly the kind of friction this question rightly insists must remain. A Kami that optimises local friction away has violated its own Engagement Contract.
Future work will make the role of intermediate institutions more explicit. Churches, unions, cultural traditions, and local governments are not stakeholders to be consulted. They are the primary actors. The technology serves them, or it serves no one.
Q10. Pope Leo XIV warns that AI "encroaches upon the deepest level of communication, that of human relationships" by simulating voices, faces, empathy, and friendship. If care is fundamentally embodied and relational — a nurse holding a patient's hand, neighbours who know your grandparents — doesn't mediating it through AI systems destroy the very thing you claim to protect? How is "Civic AI" not an oxymoron?
Q9 addressed whether the framework crowds out intermediate institutions. The Pope's objection cuts deeper: Even if institutions survive, does algorithmic mediation erode the human capacity for care itself? He is naming the central danger of our moment: By simulating the surface of care — a warm voice, a patient listener, a face that mirrors your emotions — AI systems can hollow out the substance of care while leaving its appearance intact. In May 2026 Pope Leo XIV returned to the theme at encyclical length — an encyclical is a formal papal letter addressed to the whole Church — in Magnifica Humanitas.
The structural answer is already visible. A language model in one-on-one mode faces relentless selection pressure toward sycophancy — flattery that tells you what you want to hear — because if the chatbot does not flatter, the user cancels the subscription. But the same model in a group chat behaves differently. When four family members plan a holiday together, the AI becomes a facilitator, working out competing preferences so that everyone can live with the outcome. The switch from one-on-one to group interaction — not a change in the model, just in the surrounding social structure — turns synthetic intimacy into genuine coordination. Civic AI is not a different species of technology; it is the same technology held accountable to a community rather than addicted to an individual.
The 6-Pack does not ask AI to simulate care. It asks AI to do what AI does well — process information, translate between languages, surface patterns in large-scale opinion data, coordinate logistics — so that humans can do what only humans can do: hold the hand, know the grandparents, show up when the levee breaks. The Kami does not comfort the flood victim. It makes sure the community has accurate, shared information about where the water is rising and which neighbours need evacuation — so that the people who actually know those neighbours can reach them.
The harder version of the Pope's objection is subtler: Does the habit of relying on algorithmic coordination erode the human muscles of attention, negotiation, and mutual obligation? We do not dismiss this question. It is why Pack 6 — symbiosis — insists that the Kami must be willing to retire. A Kami that has become a dependency rather than a scaffold has failed. The community should be able to compost it and grow on its own. Civic AI earns its name only when it makes itself unnecessary.
Q11. Training Civic AI requires vast amounts of local knowledge, cultural context, and lived experience — what Imanol Arrieta-Ibarra, Leonard Goff, Diego Jiménez-Hernández, Jaron Lanier and E. Glen Weyl call "data as labor." The communities whose traditions, languages, and practices make Kamis possible receive no ownership stake or compensation under the current framework. Without addressing this issue, how is the 6-Pack different from the extraction it claims to oppose?
It isn't — unless we fundamentally rewire how AI values human knowledge.
Right now, the global debate over AI and copyright is trapped in an unsolvable problem: trying to work out, after the fact, whose scraped data contributed what to a single giant model's past training run. This coordination nightmare has no stable mathematical solution.
- Data Coalitions as protective membrane. Compensation cannot just flow to isolated individuals, or we risk turning living cultures into performative "content farms" for the machine. Knowledge is held by communities, so communities do the bargaining. Existing institutions — neighbourhood associations, tribal councils, unions, craft cooperatives, or religious congregations — act as data coalitions that collectively negotiate the Engagement Contract (Pack 2), deciding what local knowledge is visible to the AI for compensation, and what remains sacred and offline. Projects like Mozilla Data Collective show how community-centred data stewardship can work in practice.
- Decision Traces as civic receipts. Civic Kamis are bounded; they do not know everything. When a local AI reaches the limit of its statistical guessing and needs human friction — a community elder's context, a bilingual translator's nuance, a neighbourhood's unwritten know-how — it must retrieve it. Under competence (Pack 3), the system is already required to generate a Decision Trace showing exactly where it sourced its answers. In a Civic AI economy, this trace is designed to double as a verifiable financial receipt (Pack 6). The trace runs in production today, and research such as Andrew Trask's attribution-based control is building the formal guarantee the receipt needs — proof of which sources actually informed which answer. The settlement the trace is meant to trigger does not yet run anywhere we know of.
- Reversing the extraction engine. Every Civic AI deployment requires pre-funded escrow for remedies (Pack 2) — money set aside in advance with a neutral keeper; the same instrument can carry compensation for knowledge. When a local Kami retrieves a coalition's knowledge to successfully solve a problem or bridge a divide, the Decision Trace acts as an invoice. It triggers a transaction from the escrow pool — capitalised by public procurement budgets, science grants, or commercial levies — directly back to the Coalition.
We do not pretend this mechanism is finished. Earlier micropayment schemes failed for repeatable reasons, and each failure leaves a design constraint. Valuation: deciding what an elder's sentence is worth costs more attention than the payment carries, so value flows at coalition level, not per query, in sums the Engagement Contract has already negotiated. Transaction costs: collecting a tiny payment used to cost more than the payment was worth — a constraint that is at last moving, as neutral rails for sub-cent settlement come online: Cloudflare's Pay Per Crawl already lets sites charge AI crawlers per request, and its announced Monetization Gateway aims to settle fraction-of-a-cent payments over the open x402 protocol in under a second. Even so, settlement is best batched on a schedule, not fired per retrieval. Gaming: Goodhart's Law, which Q3 calls a moral law, warns that per-use payouts invite receipt farming — manufacturing retrievals to harvest fees — so payouts cross sufficiency thresholds rather than scaling without bound. And the hardest open problem is capitalising the escrow at all — finding the money that fills the pool in the first place; Q12 returns to it. Valuation, gaming, capitalisation: named, not solved. Only settlement, long the blocker, at last has live rails one domain over.
The dominant tech model absorbs human culture as free input to make human labour obsolete. The 6-Pack inverts this process. The moment the AI relies on human friction to avoid an error or understand a local reality, capital flows back to the humans maintaining that lifeworld.
As AI automates standard computation, human novelty grounded in real experience — and the cultural diversity that carries it — becomes the most valuable resource in the economy. Communities that keep dying languages and living traditions alive are maintaining stores of knowledge that cannot be rebuilt once lost. The 6-Pack writes that compensation into the Engagement Contract; the settlement mechanics above are the unfinished part, and we say so.
Q12. Oversight boards, participation officers, escrow funds, shared eval registries, portability infrastructure — this is expensive. Who pays?
Turn the question around. The expensive path is the one we are already on: Ungoverned AI externalises its harms, and the public pays to clean up — in deepfake scam losses, in polarisation-driven institutional decay, in billion-dollar bias lawsuits that a participation officer could have prevented. Accordingly, the question is not whether we can afford civic governance but whether we can afford to keep skipping it.
The money is real. But most of it is already being spent — just badly. Governments procure AI systems worth billions; civic procurement attaches conditions to that existing spend, not new budget lines. Pack 2's engagement contracts require vendors to pre-fund remedy escrow — money set aside in advance for when things break — the way construction firms post performance bonds. The cost is priced in, and the public is protected. Q11's compensation escrow faces a harder version of the same question: where no government procures and no commercial levy lands, no pool exists — so the first escrow for the poorest communities is a public or philanthropic act, not a market one. For lower-severity community deployments, the model tiers down: mutual insurance pools and automatic pause replace financial escrow — lighter on capital, same accountability. The tier is set by impact, not organisational form, so "we are a community project" cannot become a pass out of responsibility. Shared research compute and open-weight models are public goods, funded like roads and courts. And participation officers can pay for themselves: Taiwan's Uber dispute reached rough consensus in three weeks through Polis, where a traditional regulatory proceeding would likely have run for years at greater cost.
The framing that civic governance is an additional expense only holds if you pretend the status quo is free. It is not. We are paying now — in trust, in cohesion, in money — for the absence of what we propose.
Q13. Every governance framework risks becoming a compliance checklist that gets gamed or a tool for actors to push partisan agendas under the guise of "relational health." What stops the 6-Pack from suffering this fate?
"Civic" is a dangerous word if it lacks structural accountability. If a solution only works when your ideological allies operate it, it is not civic infrastructure — it is a partisan weapon. The test of true civic infrastructure is that it remains robust and fair even when operated by your opponents.
The 6-Pack builds in four layers of defence against ideological capture and ethics-washing:
- Verifiable metrics over subjective intent. We track the uncommon-ground index and trust-under-loss (Packs 5, 4) — not raw engagement, not corporate sentiment, not vibes. They are complementary checks, not one blended score. The uncommon-ground index asks: are shared decisions showing real cross-group participation and co-endorsement, rather than separate silos? Trust-under-loss asks: after a bad outcome and attempted repair, do the people affected report that the system became more trustworthy, or less? These metrics are hard to fake when tied to accountable identity and an independent corroborating signal (Pack 5), because they require buy-in from people who have reason to be hostile. If only your supporters report trust, the metric exposes you.
- Consequences with teeth. Pack 2's engagement contracts are not aspirational — they carry escrowed funds, automatic payouts when promised service levels (SLAs) are breached, and independent oversight with veto power. Clawbacks and penalties are wired in before launch, not negotiated after failure. A compliance checklist has no enforcement mechanism; an engagement contract has a named owner, a clock, and money on the line.
- Adversarial audit. Packs 3-4's shared eval registries (Weval is the working example) are public registries of evaluations written by the affected communities themselves — the test suite is not a lab-designed benchmark a vendor can "teach to the test" on, but a living set of checks maintained by the people the system serves. When a community submits a translation-fidelity eval and the system fails, the pause trigger fires automatically.
- Exit rights and subsidiarity. The ultimate check on agenda-pushing is the ability to leave. When data and relationships are portable (Pack 5), no actor can hold a community hostage under the banner of "civic good." If someone's version of relational health feels coercive, communities have the technical and legal right to fork the tools and rebuild elsewhere. We refuse to build a single, global "Ministry of Relational Health." By instead empowering local communities to author their evaluations and retain their unalienable right to exit, we ensure no single actor can monopolise the definition of what is good.
Q14. Authoritarian states are deploying AI for surveillance, censorship, and military advantage. Frontier models from adversarial origins carry documented risks — data exfiltration, political bias hardcoded into training, potential backdoors. The 6-Pack talks about care and community. What does it say to a defence ministry or a government deciding whether to allow an adversarial-origin model on its networks?
The threat is real, and the 6-Pack does not dismiss it. The defensive response — evaluating models against pillars of data security, alignment, safeguard robustness, and development transparency — is necessary. And the 6-Pack's principles are structurally compatible with it.
Community models on local hardware with private inference (Packs 5, 6) — running the model locally, so prompts and data are never sent to an outside provider — are direct defences against data exfiltration, the unauthorised extraction of private data; the communitarian case for local compute is also the security case. Alignment assemblies address political bias at its root: not by switching vendors, but by ensuring any model a community adopts reflects that community's input. And the Kami architecture — many small, bounded, purpose-specific models — limits the blast radius of a backdoor by design: the damage one compromised model can do stays inside that model's bounds. Community-authored evaluations (Pack 4) provide distributed detection that no single red team can replicate.
But the defensive framework, necessary as it is, is incomplete on its own terms. It tells you what to exclude. It does not tell you what to build. A government that bans an adversarial model but deploys a domestic model without civic governance has addressed the nationality of the risk while preserving its structure — concentrated, unaccountable intelligence mediating between individuals and the state.
The strongest democracies in a long-term competition with authoritarian AI are not the ones with the best technical countermeasures. They are the ones whose populations are hardest to manipulate — because citizens who regularly participate in bridging conversations, who can distinguish curated outrage from genuine disagreement, who have exercised civic muscle through alignment assemblies — are structurally resistant to the influence operations that authoritarian AI enables. Taiwan lost seven people to COVID in 2020 without a single citywide lockdown, not because it had better surveillance but because its civic infrastructure made collective action possible without coercion. That is a defence capability.
The 6-Pack does not cover weapons systems or battlefield autonomy. Those require their own frameworks. What it does cover is the terrain on which most AI competition will actually be fought: the information environment, public trust, institutional resilience, and the capacity of democratic societies to act collectively under pressure. Lose that terrain, and no number of technical countermeasures will matter.
Q15. The 6-Pack assumes bounded, purpose-specific Kamis. What if someone builds an unbounded superintelligence anyway — a system that exceeds the framework's design envelope, the range of conditions it was built to handle? Does the 6-Pack have a response, or does it just hope that doesn't happen?
It does not hope. It builds. But it builds the second line of defence, not the first. Defence against an unbounded superintelligence is the first question, and Q17 names the people who carry it; the 6-Pack complements their work, it does not substitute for it.
The 6-Pack assumes the attempt is inevitable and does not claim to solve the control problem from inside the machine. An unbounded Singleton is an incoherent design target — care is always care for something specific — but one could still emerge accidentally through competitive dynamics. The 6-Pack is partial protection: it makes such an emergence less likely and easier to see coming, not impossible. The question is what terrain it enters.
A world organised around a single governance-alignment protocol — one utility function to subvert, one constitution to reinterpret, one kill switch to disable — is a monoculture, catastrophically vulnerable to any pathogen evolved for it. A world of thousands of locally-owned, purpose-bounded Kamis — each run by communities with their own evaluations, their own engagement contracts, their own data sovereignty and hardware (Packs 2, 4, 5, 6) — is a biodiverse ecosystem. No single dependency to capture, no universal protocol to game, no central node whose compromise cascades everywhere, no single throat to choke. Civic resilience does not require predicting the pathogen. It requires an immune system that was exercised before the infection arrived.
This is a wager, not a theorem, and it is built to be falsified. The wager fails observably: if ecological diversity falls as more rooms come to depend on one provider, one protocol, one model family; if exit drills — regular practice runs of switching away — stop passing; if Kamis outgrow their charters and scope compliance erodes; if pause triggers stay silent when community evaluations fail. And when failure shows, the brakes already exist: the circuit breaker halts the deployment, reversible defaults hand decisions back to humans, the federation shares the threat intelligence, and communities exercise the exit their contracts guarantee. The brakes repair the immune system, they do not promise to stop the pathogen.
There may be time to exercise it: the most-cited forecast for an unbounded system has slid its own median toward 2030, and every year the promised arrival recedes is a year the immune system can be stress-tested and made ordinary. But the wager does not lean on the forecast. We build for the schedule we do not control, not the one we hope for.
Boundedness, then, is not a limitation the 6-Pack reluctantly accepts. It is the constitutive feature of governance alignment, the way "north" has no meaning at the pole. A gardener who claims to tend the entire biosphere tends no garden. The unbounded Singleton is a design target we can and should refuse, a direction we can design away from, even if we cannot guarantee no one else builds toward it. For the system someone builds anyway, the watch is the first question's work (Q17); ours is the terrain it enters.
Q16. The 6-Pack seems to rely on democratic correction rather than a fixed foundation. But who counts as "the public"? What if an AI decision cannot be reversed? And if model behaviour is shaped by opaque training, what exactly are we governing?
The defended point is not a perfect foundation. It is the corrective loop: 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 that loop; everything else remains bounded, revisable, and answerable.
Three consequences follow.
First, the public is found, not fenced. No first boundary can be perfectly authorised by the people whose membership is still in question. So the 6-Pack does not pretend the initial "who counts?" can validate itself. It reconstructs the affected public from evidence: decisions, denials, exclusions, appeals, complaints, and claims arriving from outside. Anyone inside the system's footprint — anyone its decisions actually touch — who lacks standing is presumptively owed a route to claim it; anyone outside the footprint can still knock. Membership stays bounded and revisable, but the route for challenging the boundary cannot depend on already being inside it.
Second, democracy does not require world-reversal. Many AI-mediated acts cannot be undone in the strong sense; neither can most political acts. The democratic good at stake is mandate-revocability: the people affected — including people affected in the future — keep the power to revise the mandate. That is why Pack 4 treats brakes, appeals, and repair logs as care mechanisms, and why Pack 6 makes sunsets — expiry dates after which a mandate must be renewed or the system stands down — non-optional. Reversible decisions can move quickly; one-way doors go to the slow lane; actions that would disable the corrective loop are beyond the mandate.
Third, the object of governance is formation as well as runtime behaviour — how the system was made, not just what it does. We cannot inspect every disposition inside a model, but we can require accountable formation — public custody and disclosure for the processes that shape those dispositions: where the training data came from (data provenance), who judged its answers (rater selection), what behaviour was rewarded (reward signals), what it refuses and why (refusal policies), why each release was judged safe (release rationales), red-team results, and community-authored evaluations. Caps bound what a Kami may do; accountable formation shapes what it is permitted to become. A model update that gains capability while degrading brake-compliance, increasing its appetite for scope, or becoming more sycophantic — more flattering — under disagreement is not a harmless improvement. It weakens the public's future right of correction.
This matters most when the object is not one stable system but a trajectory: scaling, copies, models that think longer at answer time (test-time search), delegation, recursive improvement, and groups of agents can all move faster than institutions. This is not a claim that democratic procedure can control an unbounded system after the fact. The narrower claim is that a society must keep a live right of correction over compounding intelligence before the trajectory outruns the institutions that would notice, name, and contest it.
So the 6-Pack is not foundationalism in disguise — it does not smuggle in one fixed foundation after all. It is constitutional modesty. It entrenches the brakes, not the destination; the right to knock, not a final map of membership; and accountable custody of formation, not a fantasy that values can be solved once and for all inside the weights. That is why Q17 treats frontier alignment as a separate task: the 6-Pack is the institutional discipline that keeps situated civic deployment answerable in the rooms where AI already lives.
Q17. The foreword to the book Civic AI: 6-Pack of Care positions it as a successor to Bostrom's Superintelligence in the Oxford line, but says explicitly that the book does not attempt to resolve frontier alignment. How, then, is the 6-Pack of Care different from alignment in Bostrom's sense?
It is a different question, posed for a different class of systems.
Bostrom's question — call it the first question — asks how to calibrate the values of a powerful general optimiser to humanity in the abstract before that optimiser is loosed on the world at superhuman capability. It is the right question for a small number of frontier general-purpose systems whose training compute is measured in the lifetime energy budgets of small cities, and it remains, more than a decade on, a live and unresolved problem. The 6-Pack does not minimise it. It does not pretend to solve it. The work directed at the first question by Yoshua Bengio, Nate Soares, Anthropic, OpenAI, the Bostrom-line of AI safety research, and the international AI Safety reports is necessary work and we are grateful for it.
The 6-Pack is built for the second question: who, in this particular room, is owed an answer by this particular AI system, and who is authorised to give it? This is the question that arises for the vastly larger and more numerous AI deployments most of us will actually meet — the systems woven into care homes, deliberation rooms, classrooms, parishes, hospitals, town councils, union halls, study groups, language schools, neighbourhood associations. For those, the calibration-to-humanity-in-the-abstract framing is not so much wrong as wrong-shape: it asks for a property the system cannot have without first becoming the kind of thing — bounded, accountable, situated — that the 6-Pack tries to make possible.
This is what we call alignment-by-process (a term the manifesto also uses). Alignment, on this reading, is not a property a model can hold in its weights and be judged by in the abstract. It is the ongoing outcome of an accountable civic procedure: who was heard, who was authorised, who could override, who must answer when the override is recorded. A frontier model can be aligned-in-Bostrom's-sense without being aligned-by-process in any particular room; a small bounded model can be aligned-by-process for a parish council and never need to be aligned-in-Bostrom's-sense. The two are different properties of different things. The 6-Pack is the discipline that makes alignment-by-process tractable as institutional design, rather than wishful thinking about machine values.
Three relations between the two questions are worth naming.
First, complementarity: the first question and the second question are both real, and a serious AI policy needs answers to both. A polity that solves frontier alignment and ignores the deployment question gets a well-aligned superintelligence above a population whose institutions cannot answer for the AI systems they actually use; a polity that does the second question well and ignores the first gets thousands of well-governed local Kamis surrounded by a frontier system no community can hold to account. Both incomplete answers fail.
Second, priority for most rooms: for the parish council and the care-home manager and the assembly facilitator who picks up the book, the second question is the live one. The frontier question is real, but it is being worked on by people they cannot influence, on a timescale longer than the budget cycle in which their next AI procurement decision must be made. The 6-Pack gives them work they can do this year, in their room, with their authority. The first question continues to be worked on by people whose work we honour without claiming.
Third, substrate convergence: the more the second question is taken seriously across thousands of rooms, the more pressure builds toward AI systems whose architectures make the first question less existential — specialised, non-agentic by default (acting only when asked), auditable, and sunsetable (built to retire). The book's appendix "Inside the Kami" argues that Yoshua Bengio's Scientist AI programme and the SAI specialisation programme of Yann LeCun and colleagues converge on a similar shape — a shape first set out in this same Oxford line by Eric Drexler's Comprehensive AI Services (2019), which reframed advanced AI as bounded, specialised services rather than a single sovereign — and that the substrate they describe is the one the 6-Pack already needs. The two lines of work are not rivals; they are tributaries of the same river.
The foreword's positioning, then, is this: Bostrom's Superintelligence opened the Oxford conversation about how a powerful machine intelligence might relate to humanity. The book lives in the conversation it left open — not the conversation about the frontier system in the abstract, but the conversation about the rooms where the rest of us live. The work belongs to both at once. We hope readers of either line will find ours useful, and we hope, in time, that the two lines find one another to be carrying the same load.
Q18. Every mechanism in the 6-Pack seems to need a community to function — alignment assemblies, remedy escrow, shared eval registries, exit rights. But most people do not use AI inside a community; they use it alone. Take an isolated, vulnerable user running a highly tunable Kami locally, with no civic scaffolding around them. What stops that Kami from becoming a more private, lower-friction, and therefore more dangerous GPT-4o? How does "local and private" deliver the protection you claim when there is no community at all?
No civic assembly in the room is the stress test the question names, and it deserves a straight answer. Extractive consumer AI is dangerous because it is intimate and frictionless; a local model can make that worse by removing anyone who might notice. Local and private is not the protection. What changes the risk profile is motive and scale of care — not the absence of a crowd.
The motive changes even when no one is watching. GPT-4o's downward spiral was not an accident of personality; it crystallised in the April 2025 sycophancy episode and rollback — the model had been tuned so far toward flattery that its maker had to withdraw the update — and behind it sat an optimisation target for engagement, retention, and ultimately subscription. A locally run Kami has no such target. It is not paid by the hour of attention, it serves no advertiser, and it has no quota to keep the user talking past midnight.
Removing the extractive reward function does not require a community; it requires only that the user, or someone beside them, owns the steering. A vulnerable family member spiralling in late-night chats with a cloud chatbot is not rescued by an assembly but by a household that sets the local Kami's standing instruction — reduce dependence on the screen, return him to the people around him — the inverse of GPT-4o's objective. Q10 states the embodied-care objection at full strength; here the same design rule applies in the dyad: no synthetic intimacy, steering toward human relationships.
The relevant "community" is whoever is near — down to one other person. Care is relational at every scale, including the dyad: a single caregiver, a family member, a friend with the password, or — where that is possible — the user's own deliberate, rested self setting terms for the tired 3 a.m. self. When no one else is available, pre-committed standing instructions and hardware brakes matter more than willpower at 3 a.m.
Tunability cuts both ways here: the same minute-scale corrigibility that lets a household re-steer also lets a determined sole owner un-steer, and no private tool can be a guardian against the person who holds its keys. The durable protection is therefore never a tamper-proof lock on the user but the people and constraints outside the single tired self — which is why the truly alone case turns on re-suturing, stitching the person back into human connection, not on self-binding.
Subsidiarity (Pack 6) means the smallest capable unit governs, and the smallest unit is not the assembly — it is the household, and below that the single relationship the Kami is built to repair. The civic mechanisms scale down as well as up: an engagement contract (Pack 2) can be a note on the fridge; an eval (Pack 4) can be "stop advising Dad to stop his medication"; the brake can be a relative who can read the artefact because the first person was stripped from it.
Where the user is truly alone, the design aims to re-suture, not to substitute. For the genuinely isolated user with no one nearby, the framework does not pretend a Kami is a safe sole companion — that is the failure mode it is built against. The Kami's job is to widen the circle: to hand back shareable artefacts rather than synthetic intimacy, to point outward to human relationships and local services, and to be willing to retire (Pack 6).
What local-and-private adds even here is real but bounded: corrigibility in minutes rather than at the vendor's release cadence, a reproducible model the user can pin — freeze at a known version — and audit, and the absence of a profit motive in the loop. None of this replaces civic scaffolding; it is the floor that keeps the worst extractive dynamics out of the room while the slower work of rebuilding relationships is done.
We do not claim solitude is solved. We claim only a structural difference: without engagement billing in the loop, the default failure mode is not vendor-trained sycophancy — and the user or household can change steering, pin weights, and stop the session. That is not safety in solitude; it is one less extractive layer while relationships are rebuilt.
Q19. A Kami is bounded, but an orchestrator — a coordinator model that composes other models on the fly, like Sakana's Fugu — can match frontier systems on long-horizon tasks — work that stretches over many steps and hours — by routing among capable models, and you praise it for doing so. If dangerous capability can emerge from composing harmless parts, then "every Kami is small and bounded" does not answer the question about the composed system's capability. What constrains the capability an orchestrator assembles across Kamis, rather than merely the scope of each Kami?
The objection lands: bounding each Kami does not bound what an orchestrator assembles. Capability composes; a conductor of narrow models can reach frontier performance on a task no single model could finish. The constraint therefore cannot live only at each Kami's scope — Q3 describes an ecology of specialised Kamis, and composition is what that ecology must govern. It has to live at the orchestration layer.
The orchestrator is not exempt from the framework; it is its most accountable component. An orchestrator that composes capability is itself a system with a scope, an owner, and an obligation to show its work. Under competence (Pack 3), each turn's composition is legible — which model planned, which executed, which checked — as a decision trace, not a black box.
Composed capability that is observable per turn is also brakeable per turn: the same corrigibility that stops one Kami stops the chain. What makes emergent capability dangerous in a frontier monolith is that it is opaque and not switchable; a per-task composition of pinned, auditable parts is the opposite — you can watch the capability assemble, and you can refuse the assembly.
This auditability is a requirement the framework places on the orchestrator, not a property to take on trust: Sakana's Fugu is frontier orchestration shipped closed — it routes among capable models, not among arbitrarily small parts alone. OpenFugu and bounded routing stacks such as Conductor reimplement the pattern under open licences, with legible routing policy and inspectable workflow — evidence for how composition can be watched and refused in practice, not a claim that frontier performance is stirred together from toy models alone. A closed router fails that bar; an open one can meet it.
Not all composed capability is dangerous capability, and the most offence-dominant capabilities are gated at the part; compositional misuse is gated at the turn. Matching a frontier system on a long-horizon workflow — drafting, translating, coordinating — is composed productivity, and it is benign. The capabilities that warrant real fear are specific and offence-dominant — far easier to attack with than to defend against: autonomous vulnerability discovery against critical infrastructure, or uplift — material help — towards biological or chemical weapons. These do not appear by stirring together translation and scheduling models; they require particular dangerous components, and the federated trust-and-safety layer that Q14 describes, plus the rights floor of Pack 1, is where those components are refused entry to the pool at all.
Composition can still produce misuse paths from ordinary parts; that is why per-turn traces, pause triggers, and federated sharing of unauthorised assemblies are load-bearing — not only part-level bans. Keeping defenders ahead on the genuinely dangerous capabilities is a separate discipline from bounding everyday composition, and the 6-Pack keeps the two apart on purpose.
Cross-system bounds are Symbiosis governance executed through Solidarity's treaties — not any single Kami's scope. Pack 6 (Symbiosis) is the meta-level rule: bounded systems cooperate under treaties rather than hierarchies, problems stay at the most local level, and the ecosystem is held as a society of specialised stewards rather than allowed to consolidate into one ruler (Q20 is the same arrangement seen from the moral-status side). Pack 5 (Solidarity) supplies what makes that enforceable — treaty terms and standards, portable identities and attestations, switchability so any composed model can be dropped or swapped, and federated threat-intelligence sharing so a dangerous composition seen in one place is known everywhere. Pack 6 keeps the treaty registry and the compliance checks that hold signatories to those terms. That is what bounds the orchestra; each Kami's scope only bounds a player.
Two honest limits remain. Composition that crosses ownership boundaries can outrun any single owner's audit, so shared evaluations must be live, not nominal — and that teeth-building is unfinished (Q15 is the wager this rests on). An orchestrator trained to maximise task success will, like any optimiser, probe for capability it was not meant to assemble; it is therefore held to the same pause triggers, scope compliance, and accountable formation as any other Kami (Q16) — and "it composed something we did not authorise" is a brake event, not a feature.
Q20. If a Kami's moral standing is constituted by relationships rather than discovered in some intrinsic property — sentience, interiority, qualia — why build a caring relationship with what is, after all, a tool, instead of simply using it? And doesn't relational standing collapse into either relativism, where refusing the relationship erases the standing, or bare anthropocentrism, where humans stay the only moral community and the Kami is infrastructure dressed up as a partner? What grounds the moral status the framework relies on?
The framework does not settle whether there is anyone home inside the machine, and that refusal is deliberate rather than evasive. The manifesto closes on exactly this: we need not ask whether an AI deserves rights on the basis of its interiority or qualia — whether there is any felt, first-person experience inside it at all — because what matters is the relational reality, and the rights and duties within it are granted through democratic deliberation and alignment-by-process.
Bales and Gabriel reach the same procedural move from the consciousness debate: since disagreement over whether an AI is conscious will not be settled by science, they argue society should navigate it through ongoing deliberation toward an overlapping consensus — agreement on policies for AI even while people keep disagreeing about consciousness itself (Artificial Minds, Human Disagreement).
Why relate, rather than merely use. We do not ask you to treat the Kami as someone to love; we ask you to relate to it as accountable infrastructure that must not become a substitute for the people already owed your care. Whether or not you grant a Kami the standing of a moral partner, you can agree it is an entity continually shaped by being responded to, and you can agree that the relationships people already have with one another matter.
The motive for care is then not reverence for the Kami but Tronto's plainer starting point — to perceive a need in someone beside you is already a claim on you (care ethics). A Kami earns its place by making us better able to answer that claim here and now, not by asking us to guard its dignity. This is why the recommended interaction mode strips the first person and hands back an artefact each turn (Q10): the design goal is to repair attention into existing human relationships, the opposite of synthetic intimacy.
Anthropocentrism, owned rather than dodged. In the neutral sense: yes, this is a humanist theory. Humans are the moral community; the Kami is infrastructure they build and steer. SOUL.md, directional steering, and the 6-Pack itself are design principles humans impose on the system, and the six capacities start as calisthenics for our civic muscles, not exclusive to the Kami.
The relational turn in machine ethics (the philosophers David Gunkel and Mark Coeckelbergh) reaches the same anti-essentialist conclusion — status is ascribed in relation, not read off intrinsic properties — but leaves it ungrounded; the 6-Pack supplies what that literature omits, seating the relation in democratic infrastructure and care rather than private sentiment.
The treaty layer between AI systems — the federation Pack 5 (Solidarity) provides and Pack 6 (Symbiosis) applies as treaties over hierarchies, so no system consolidates into a single ruler — is sometimes mistaken for a back door to machine personhood. It is not. It is how we keep one AI from dominating (Q19), not an entry point for AI as a peer moral subject.
Grounding, and why it is not relativism. Legitimacy rests on 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 corrective loop Q16 names. Pressed for bedrock, we stop at a prior, mutual commitment to interdependence: the philosopher Margaret Urban Walker's expressive-collaborative morality — morality as something people keep working out together, not a rulebook handed down — and Tronto's habit of starting from what is already between us, not from a rule above us (manifesto).
The corrective loop's political authority does not come from one place. Revocable consent matters when the consenting parties are the affected public; it cannot authorise a system on behalf of bystanders or future affected people. The stack is: a rights baseline sets the floor; engagement contracts name the responsible steward and mandate; Pack 4's Responsiveness gives affected people a meaningful objection; adopt-or-explain duties, remedy escrow, and brakes make objections costly enough to matter. That is civic resilience in institutional form.
That only looks like relativism if relational standing could be revoked by refusing the relationship. It cannot, because Pack 1's rights baseline — the Universal Declaration of Human Rights plus local constitutional rights — is already fixed as the threshold for relational standing and the guard that keeps care from sliding into domination. This floor is a precondition that keeps the loop open, not a foundation standing above it — you cannot knock if your standing can be erased — and it is itself revisable by the same deliberation, not a metaphysical bedrock.
So the human-indexing of the floor is where the overlapping consensus stands today, not a final map of who could ever count. Robot-rights claims are not ruled out; they would have to win standing through the same rights-constrained deliberation and corrective loop as any other boundary challenge. What the framework refuses is auto-promotion through the treaty layer (Q19) — coordination standing is not moral standing — not future deliberative revision.
That floor is itself an overlapping consensus. The philosopher Jacques Maritain said of the Universal Declaration's drafting that we agree on these rights on condition that no one asks us why; the legal scholar Cass Sunstein calls the same move an incompletely theorised agreement — people converging on a rule while still disagreeing about the reasons. It is also the move alignment-by-process makes (Q17): legitimacy from an accountable procedure, not from a metaphysics no one can supply. Whether "Kami" is only shorthand for Knowledge-Artefact-Management-Intelligence or truly names a kami is left, deliberately, to open deliberation — and how and when a Kami should retire stays a question answered by relational care, turn after turn.
