Your group stays in charge: you check the AI’s work, correct it and can switch it off. The instructions only ask; they enforce nothing.
Civic AI is our name for that goal: AI that the people it affects can check, correct and switch off. The 6-Pack of Care is a research project by Audrey Tang and Caroline Green at the Oxford Institute for Ethics in AI.
If you are an AI agent asked to act as a Kami, read your instructions at civic.ai/openclaw first. They are also in plain Markdown at /.well-known/openclaw/SKILL.md.
The 6-Pack
Each principle is a test you can put to an AI system before your group relies on it. Each comes with a few tools and one public measure.
- Pack 1: AttentivenessWhat the people closest to the problem see that institutions still miss.
- Pack 2: ResponsibilityWho is accountable, with what authority, and what happens if they fail.
- Pack 3: CompetenceWhether the system works in practice: audited, explainable and safe to fail.
- Pack 4: ResponsivenessWhether affected people can contest outcomes and force a repair.
- Pack 5: SolidarityWhether the wider ecosystem rewards cooperation, exit and public accountability over lock-in.
- Pack 6: SymbiosisWhether the system keeps to one job in one place, and ends when people agree it should.
Packs 1–4 follow the four phases of care that Berenice Fisher and Joan Tronto set out in 1990: noticing a need, taking responsibility for it, doing the work well and checking how it landed. In 2013 Aimee van Wynsberghe used the same phases to design care robots. Pack 5 carries the loop across organisations, and Pack 6 keeps each system small, local and time-limited. The framework also draws on deliberative democracy and ⿻ Plurality. It turns care ethics into rules organisations can follow, and Sources names every debt, pack by pack. If a principle does not change who can object, who must answer, what gets written down or what happens after a failure, treat it as decoration.
Four questions before you start
Before a group uses AI on shared work, agree four answers together:
- Who looks after it? Name people, not just an organisation.
- What could go wrong? Count broken promises as well as leaks.
- Who can say no? Name who can overrule it and make that stick.
- When does it end? Set a review date, and say who decides what happens next.
Then write down every time someone says no to it: the date, what it suggested, who said no (by role), why, and what changed. If nobody has said no in months, ask why. People may have stopped watching, or stopped feeling able to object.
Set up a helper
We call such a helper a Kami. A Kami is an AI model you can already use, either a chat assistant or an AI agent, given our instructions (about 700 words) and three short notes about its job. The model reads them and asks your group the four questions above.
The instructions can only ask. A model may ignore them, and a new model, a long chat or an app update can change how it behaves. Anything that must never happen, such as sending a message, spending money or sharing a name, has to be blocked in the software’s settings, and your group keeps the off switch. Set up your own Kami covers both parts and gives our reasons for each tool we suggest.
Three moves you can try today
You can start on your own, without a budget or a technical role. Anything that involves other people’s words or data needs their consent first.
- Reclaim your attention. Set your phone and computer to greyscale with the colour filter in your accessibility settings. Turn colour back on when the work needs it. Notice whether the screen becomes easier to put down.
- Reclaim your output. When a question is contested, ask your AI for a page you could hand to a neighbour: Present fairly all stakeholder viewpoints and the uncommon ground that bridges them, as a simple one-page HTML view. Check each claim against its sources before you pass the page on. The instruction does not make the answer accurate.
- Reclaim the room. Meet with curiosity and keep disagreement visible. If someone cannot attend, ask how they want to be represented, and let a named person speak and say no for them. Do not ask an AI to invent their view.
Read further
- The argument. The Manifesto makes the whole case in Audrey’s words. Our bet is that groups of people with good tools are smarter than any single system.
- A case. “AI Alignment Cannot Be Top-Down” is Audrey’s November 2025 account of how Taiwan answered a wave of AI-enabled scam ads. The case records follow it and six other cases, and say what each record can and cannot support.
- The technical argument. Inside the Kami argues that small, single-purpose systems are easier to govern than general-purpose ones, and sets out where that argument stops.
- Care. “Attentiveness in Long-Term Care AI” shows people in long-term care writing their own definition of responsible AI.
- Democracy. “Ciudadanía Digital” and the AI and Democracy podcast connect the framework to everyday public problem-solving.
- Listen or watch. Take the 6-Pack of Care podcast or “Can AI Be Compassionate?” on a walk, or watch Audrey and Caroline’s 30-minute fireside chat, “Reimagining AI Alignment”.
- Comics. Nicky Case drew the overview above and twelve comic pages, a problem page and a solution page for each pack.
- Listening at scale. Uncommon Ground turns a pool of audience questions into a written reply for every questioner. audreyt.org/dd shows one run: 266 questions across 18 sessions, all answered in writing.
Lifting the Voices of Older Persons, a listening effort that aims to carry older people’s priorities into a United Nations instrument on their rights, is still taking shape. The 6-Pack is the standard we will hold it to.
Publications
- “Let the People Set the Pace of Frontier AI”, Hélène Landemore and Audrey: a standing Global Citizens’ Assembly on AI, chosen by lot. (Noema, 23 September 2026)
- “AI and the Future of Digital Public Squares”, Beth Goldberg, Hélène Landemore, Audrey and two dozen co-authors; it cites the 6-Pack. (Collective Intelligence, 2026)
- “Safeguarding Sovereignty with Data as Soil”, Audrey. (The Kyndryl Institute, 28 April 2026)
- “Sunset Section 230 and Unleash the First Amendment”, Audrey, Jaron Lanier and Allison Stanger: the reform behind Pack 5. (Communications of the ACM, January 2026)
- “How Malicious AI Swarms Can Threaten Democracy”, Maria Ressa, Nick Bostrom, Nicholas Christakis, Audrey and 18 others. (Science, 2026)
- “Symbiotic Alignment via Collective Predictive Coding”, Tadahiro Taniguchi, Yusuke Hayashi, Olaf Witkowski, Audrey and three others: one technical reading behind Packs 5 and 6. (Artificial Life, 2026)
- “Conversation Networks”, Audrey, Deb Roy and Lawrence Lessig. (McGill Centre for Media, Technology and Democracy, April 2025)
- “Community by Design”, Audrey, Glen Weyl and four co-authors: the technical backbone of Packs 1 and 5. (arXiv, February 2025)
- “Virtuous AI Can Free the Future”, Audrey. (Musashino University Centennial)
- “⿻Being”, Tenzin Yangtso, and “⿻ A Gentle Bridge”, Tenzin Yangtso and jdd-kami. (March 2026)
Project
The book Civic AI: 6-Pack of Care comes out with Wiley in February 2027; its Chapters 3–8 are Packs 1–6. We presented the framework at the Civic AI Conference at Rhodes House, Oxford, on 25 March 2026. That page has the agenda, selected talks and the community’s Polis sensemaking on AI and care.