The 6-Pack of Care is not merely theoretical. Throughout this book, we have drawn on real-world examples to illustrate each pack in turn — the jolly hostess of Attentiveness, the signed work order of Responsibility, the workshop wall of retired broken parts that Responsiveness builds. In this chapter, we bring several of those examples together, and add new ones, to show how civic AI operates as a coherent system across different scales, cultures, and problem domains.
What unites the cases we have studied is not a shared ideology or a common technology stack. It is a shared orientation: care as the organising principle of design. Before a system is built, someone has asked: who is affected? Before it is deployed, someone has asked: who can stop it if it goes wrong? And after it has run, someone has asked: what did the community learn, and what can the next steward inherit?
These are not natural questions for technology projects. They require a different set of instincts — what we have called civic care. The approach is relational: it brings people together around shared problems rather than optimising for individual outcomes. It is inclusive: not limited to technical experts, accessible regardless of geography or political standing, built with and for the people it affects. And it is accessible: the methods and frameworks described in this book are available — not locked behind proprietary walls or academic paywalls, but offered as tools anyone can pick up, adapt, and govern locally.
This chapter explores five cases: a national assembly process in Taiwan, a wildfire recovery platform in California, an electoral experiment turned political movement in Japan, a language preservation project for Tibetan communities, and a federated child-safety network operating across platforms on four continents. Together, they trace the full range of what civic AI looks like in practice.
Case Study 1: Taiwan's Alignment Assemblies
Chapter 1 told this story in full: how 447 randomly selected Taiwanese citizens, deliberating in groups of ten with AI-assisted sensemaking, produced a policy framework for deepfake scam prevention that achieved 85 per cent consensus, was enacted by parliament within months, and reduced deepfake scam incidents by 94 per cent — all without resorting to censorship. Here we revisit the case to show how the 6-Pack operates as a coherent system rather than a collection of isolated principles.
Looking at this case through the 6-Pack, the mapping is direct. Attentiveness operated through broad random sampling: the mini-public was not drawn from existing civic associations or expert registers, but from the population at large, including people who had never previously participated in policy deliberation. The AI sensemaking tools were explicitly designed to surface what we might call the uncommon ground — not the loudest consensus, but the proposals that earned cross-group endorsement. Responsibility was built into the process architecture: the pre-committed engagement contract meant that the government had posted its work order before a single participant arrived.
Competence was visible in the AI assistants' role. They did not propose solutions. They provided real-time transcripts, made visible where summaries diverged from source contributions, and gave participants the information they needed to deliberate rather than the conclusions they should reach. Responsiveness was structural: the 85 per cent consensus figure was published alongside the dissenting 15 per cent, ensuring that the outcome could be contested and scrutinised. Solidarity was achieved not through top-down mandate but through a process that required cross-partisan support — the assembly's design made it mathematically impossible for any single faction to impose its preferences without genuine persuasion. And Symbiosis was honoured in the assembly's boundedness: it had a specific purpose, a specific duration, and it ended when its work was done. The Kami retired; the law remained.
Taiwan's assembly also demonstrates that deliberation is not inherently slow. The process moved from convening to enacted legislation in months — faster than most corporate policy cycles. The speed was possible because the assembly was not an event added onto a pre-existing policy process, but a piece of standing civic infrastructure, with rules and authorities already established. As we argued in Chapter 9, deliberation is slow only when it is treated as an exception rather than a foundation.
Case Study 2: Engaged California
On the night of 7 January 2025, wildfires ignited in Los Angeles County. Within days, the communities of Altadena and Pacific Palisades were facing catastrophic destruction. Thousands of homes were lost. The recovery question — not just how to rebuild, but what to rebuild, and for whom, and by what process — was among the most politically contested the state had faced in decades.
Governor Gavin Newsom's administration launched the Engaged California platform at engaged.ca.gov with a mission statement that was also a design commitment: to co-create recovery with the community. This was not the language of public consultation, which typically means gathering feedback on decisions already made. It was the language of civic care: the community is not an audience for governance, but a participant in it.
The platform worked with approximately 900 people from Altadena and Pacific Palisades — survivors, community organisers, small business owners, renters and homeowners, long-time residents and recent arrivals. It did not select for technical expertise or civic engagement history. It selected for proximity to the harm. This is precisely what Attentiveness demands: not the most articulate voices, but the most affected ones.
AI sensemaking tools — similar in architecture to the tools used in Taiwan's assemblies — helped facilitators identify where recovery priorities converged and where they diverged. The outputs were not averages. They were maps: here is where almost everyone agrees; here is where communities split along lines of tenure and income; here is the proposal that earns endorsement across those lines. These brochures, as we describe them in Chapter 3, were shareable, citable, and correctable.
Meanwhile, a parallel strand of the platform engaged more than 1,400 state employees, who submitted over 2,600 ideas on government efficiency in wildfire response and recovery. These were not suggestions dropped into a void. They informed real executive action — changes to procurement, permitting, and inter-agency coordination that directly affected the pace of recovery on the ground. Responsiveness here was structural: the platform was designed so that ideas had a traceable path from submission to outcome, or a published explanation of why they had not been adopted.
The solidarity dimension of Engaged California is perhaps the most significant for the book's argument about scale. Recovery after a disaster is profoundly fractured: different communities have different losses, different resources, different timelines, and different political relationships to the institutions governing them. Altadena, a historically Black and multiracial community, had a different experience of the fire and of local government than Pacific Palisades. Engaged California did not flatten this difference. It made it legible — and in doing so, created conditions for solidarity that were grounded in recognition rather than false unity.
What distinguishes Engaged California from a well-intentioned consultation exercise is the trajectory toward permanence. A bill working through the California legislature in 2025 proposed making the platform into standing civic infrastructure — not one governor's initiative, dependent on one administration's commitment, but a durable institutional resource available to any future government facing any future crisis. This is the institutional memory that Responsibility requires: not a one-off promise, but an accountable process with teeth that survives leadership change.
If the legislation succeeds, California will have done something that democratic theory has long aspired to and rarely achieved: created a standing capacity for participatory governance that meets people at the scale of their actual lives — local, specific, and responsive — rather than at the scale of the state's administrative convenience.
Case Study 3: Team Mirai and Electoral Civic AI in Japan
In the summer of 2024, a software engineer named Takahiro Anno read Plurality — the book co-authored by Glen Weyl and Audrey Tang on technology and democracy — and decided to run for governor of Tokyo. His campaign was, by any conventional measure, low-budget and politically marginal. But the way he ran it was not conventional.
Anno campaigned partly as a VTuber — a virtual YouTuber, using a digital avatar and livestreamed content to reach audiences that traditional political campaigns do not reach. More significantly, he built a process by which anyone could call his AI assistant and suggest improvements to his platform. Those suggestions were aggregated, surfaced, and announced publicly on YouTube. The AI was not making policy decisions. It was doing what the 6-Pack calls broad listening: taking input from distributed sources, keeping it legible, and returning it to the community as a shared artefact rather than a private data asset.
Anno received more than two per cent of the vote. He did not win. But the campaign was not simply an electoral experiment — it was a proof of concept for what civic AI can do in a context as traditionally resistant to democratic innovation as electoral politics. The signal was narrow but true: it was possible to build a participatory political platform at low cost, at scale, without pretending that the AI was making decisions it was not making.
Anno did not return to software engineering after the election. He founded Team Mirai — translated as the Future Party. In subsequent elections, Team Mirai won eleven seats in Japan's lower house, and Anno himself became a member of the upper house. The party is now bringing the civic AI conversation into the Japanese Diet — not as a technology policy question, but as a question about the relationship between democracy and the tools that mediate it.
The lesson here is not that civic AI can win elections. It is subtler. Even the most entrenched and symbol-laden democratic institutions — the ballot, the campaign, the legislative chamber — are not immune to reimagining through civic AI infrastructure. Anno's campaign demonstrated that a candidate can build genuine participation into their platform without either faking it (pre-determined outcomes dressed as consultation) or abandoning it (the candidate decides alone, participation is theatre). The AI was a tool for genuine attentiveness, and it was visible enough that voters could evaluate whether it was working.
Team Mirai also demonstrates the 6-Pack's principle that boundedness is not a limitation. Anno's AI assistant was not trying to govern Japan. It was trying to listen to the people who wanted to contribute to a gubernatorial platform in Tokyo. The narrowness of its purpose was the source of its credibility. It was a Kami of the campaign — tending its specific ground, accountable to the people who used it, ending when its purpose ended.
Case Study 4: Monlam AI and the Preservation of Tibetan
Not all civic AI operates at the scale of national governance. Some of the most important work happens in bounded cultural communities, where the stakes are the survival of a language and the lifeworld it carries.
Monlam AI is a project led by a small team of Tibetan computer scientists, anchored by the scholar and technologist Geshe Lobsang. Over years of patient work, the team has digitised thousands of Tibetan texts — classical manuscripts, religious documents, scholarly commentaries, oral tradition transcriptions — and built a suite of AI tools including instant translation between Tibetan and other languages, a digital dictionary, and applications that are accessible to Tibetan communities worldwide, including communities in diaspora who have no access to physical archives.
Tibetan is a language under severe pressure. Its speakers are geographically dispersed, its formal institutions are constrained, and its textual tradition is vast and partly inaccessible even to many native speakers. The kind of AI that Monlam builds is not trying to replace the scholars and monastics who hold this tradition — it is trying to give them tools that match the scale of the problem they face. A scholar who can search ten thousand texts in seconds can do work that was previously impossible in a single lifetime.
This is Symbiosis in its clearest expression. The Monlam AI is bounded to a specific cultural need. It was not built to be everything for everybody; it was built to do this specific thing, for this specific community, with exceptional care for accuracy and cultural fidelity. The governance of the project lives with the community it serves: the tools are developed in consultation with the people who use them, and the decisions about what to digitise, how to render ambiguous passages, and which translations to privilege are made by people with deep cultural authority, not by external algorithms trained on majority-language data.
The contrast with extractive AI is instructive. A large general-purpose language model trained on web-scraped data will have encountered Tibetan text, but its representation of Tibetan culture will be thin, distorted by what was digitally legible, and ungoverned by the community whose tradition it claims to represent. Monlam's approach is the opposite: deep, specific, community-governed, and designed to preserve rather than extract.
There is also a solidarity dimension to the Monlam case that deserves attention. The project makes tools accessible to Tibetan speakers worldwide — in India, in Europe, in North America, in communities that have been in diaspora for generations. It builds a kind of distributed cultural infrastructure that does not depend on physical proximity to a single archive or institution. In the language of Chapter 7, this is what federation looks like when the shared resource is not threat intelligence but cultural memory: local in governance, global in reach, and structured to benefit the community that holds the knowledge rather than the platforms that host the tools.
The Monlam case also speaks directly to the question raised in Chapter 9 about data as labour. The texts that Monlam has digitised represent centuries of intellectual and spiritual work by Tibetan scholars, monastics, and scribes. The AI tools built from those texts derive their value from that labour. Monlam's structure ensures that the value flows back to the community — not as a charitable gesture, but as a structural feature of the project's governance. The community owns the tools, governs their development, and determines their use.
Case Study 5: ROOST and Federated Trust and Safety
Child sexual abuse material — CSAM — is among the most serious harms the internet enables and among the most technically difficult to address. The difficulty is not only technical. It is structural. Effective detection requires training on examples of abuse, which raises profound ethical and legal questions. It requires coordination across platforms, which raises questions about data sharing and privacy. And it requires enforcement at scale, which has historically meant either delegating enormous power to a small number of centralised services or accepting that significant harms will go unaddressed.
ROOST — Robust Open Online Safety Tools — offers a different architecture. Rather than routing all content through a central detection service, ROOST trains local models that run on standard hardware, including laptops, under a community code of conduct. Each incoming piece of content receives not a binary determination but a set of citations — references to the specific patterns or examples that triggered a flag — that can be contested, reviewed, and appealed by community moderators. An audit trail records every decision.
This is the federated approach described in Chapter 7 made concrete for the hardest possible problem. ROOST does not require platform operators to send private user content to a central service. It does not require them to trust a single vendor's definitions of harm. It allows them to apply shared safety intelligence while retaining local enforcement authority — and it gives their moderators the tools to contest and improve the model's judgements over time.
By 2025, ROOST was in production use across Discord, Bluesky, Roblox, and Notion — platforms serving very different communities, with very different community norms and moderation cultures, but sharing a common interest in effective child protection that does not require sacrificing the privacy of the overwhelming majority of their users who are not perpetrators.
The Solidarity mapping here is direct. ROOST embodies what Chapter 7 calls the federated safety network: partners detecting harms in their own cultural context, sharing threat signals, and keeping enforcement local rather than ceding it to a single hub. The open-stack architecture — open weights, open audit trails, open community code of conduct — means that no single actor can monopolise the definition of what counts as harmful or capture the governance of the tools for narrow commercial purposes.
There is a broader implication. ROOST demonstrates that the tension between privacy and safety, which is often presented as a forced choice — you must either scan everything centrally or accept that harms go undetected — is a false binary. The federated architecture resolves the tension not by choosing one value over the other but by redesigning the system so that both can be honoured. Local processing preserves privacy; shared signal intelligence preserves safety; community governance preserves accountability. None of these is sacrificed to the others.
This is also a case where Competence and Responsiveness are load-bearing in practice. The contestable citations give moderators a meaningful right of reply — not an appeal to a black box, but a legible challenge to a specific, auditable decision. When a citation is successfully contested, the outcome is recorded and can update the shared model. The wall of retired broken parts is not just a metaphor here; it is the mechanism by which a federated safety network improves over time without requiring any single platform to cede control of its moderation to an external authority.
Conclusion: The Pattern Behind the Practice
These five cases span different scales — from a community of Tibetan scholars to a national legislative process — and different cultures: Taiwan, Japan, the United States, Tibet, and a global network of online platforms. They address different problem domains: financial fraud, disaster recovery, electoral politics, cultural preservation, and child safety. No single technology underpins all of them. No single organisation coordinates them.
What unites them is a pattern.
Each case begins with local attention — not a global ambition to fix everything, but a specific question asked by a specific community about a specific harm. Taiwan asked how to protect citizens from deepfake scams without resorting to censorship. Tibetan scholars asked how to make a vast textual tradition accessible to a dispersed community. California wildfire survivors asked how to shape their own recovery. In each case, the starting point was a relationship — between people and the problems affecting their lives — rather than a technology looking for applications.
Each case embeds shared accountability — structures that make it possible to verify whether commitments are being kept and to contest outcomes when they are not. Taiwan's pre-committed engagement contract, Engaged California's traceable idea-to-action pathway, ROOST's contestable citations: these are all different implementations of the same underlying requirement. Care without accountability is sentiment; accountability without care is compliance theatre. The 6-Pack insists on both.
Each case respects bounded purpose — the Kami principle that a system governs what it is for and nothing more. Monlam AI tends Tibetan language and culture; it is not trying to govern Tibetan politics. Anno's AI assistant listened to gubernatorial platform suggestions; it was not making policy decisions. ROOST detects CSAM; it does not adjudicate the broader content policies of the platforms that use it. This boundedness is not a limitation — it is the source of each system's trustworthiness.
And each case enacts community governance — the principle that the people closest to the problem have authority over the tools designed to address it. Not merely a seat at the table, but the ability to write their own evaluations, contest decisions, and if necessary, shut the system down. This is what distinguishes civic AI from institutional AI that is deployed on communities rather than with them.
The 6-Pack of Care is not a template to impose on these situations. It is a lens through which to recognise what they have in common — to see, in the assembly rooms in Taiwan and the mountain monasteries in Tibetan communities and the trust-and-safety teams on platforms serving millions of children, the same underlying structure of care: attentive to the specific people affected, responsible to verifiable commitments, competent in execution, responsive to failure, solidaristic in sharing the costs of governance, and symbiotic in treating the system's own retirement as a sign of success.
These cases also suggest where the framework still has work to do. Each of them emerged in contexts with existing civic infrastructure — a functioning legislature, a responsive governor's office, a scholarly community with institutional continuity. The harder tests lie in contexts where that infrastructure is weaker: where civic participation has been actively suppressed, where communities are dispersed or persecuted, where the state is more likely to be a source of harm than a partner in governance. The 6-Pack names the principles. Building the institutions that give those principles teeth in harder soil is the next generation of work.
For now, these cases offer something more modest but genuinely valuable: proof that civic AI is not science fiction. It is already happening. The assembly in Taiwan, the platform in California, the senator in Tokyo, the dictionary in Dharamsala, the federated safety network running on a laptop in a moderator's home — these are not prototypes awaiting deployment. They are deployments, tested under real conditions, with real consequences for real people.
The question the next decade will answer is not whether civic AI is possible. It is whether we choose to build it at scale — and whether the communities who need it most will have the standing to shape it.