What is the 6-Pack of Care?

Six principles for checking whether AI actually serves people — not only by looking inside the AI, but by asking whether people on the ground were heard, who is accountable, whether failures can be repaired, and whether it can be switched off. Audrey Tang and Caroline Green set them out in 2025.

As AI comes into use in government offices, hospitals and care homes, a question arises: who checks whether this AI is really working, and how? The usual assumption is that if the company building the AI writes good rules and trains it well, we can relax. The 6-Pack of Care holds that this is not enough. It sets out, in six parts, procedures that let the people who work with an AI check it, repair it and stop it themselves.

In short

The 6-Pack of Care is a set of six principles for AI governance (the rules for how AI is run and who is responsible for it). It was published in 2025 at the University of Oxford by Audrey Tang, Taiwan's former Digital Minister, and Caroline Green [1][2].

It asks about more than the AI's performance. In the words of the project's website, "It asks who gets heard, who is accountable, how failures are repaired, and when a system should stop" [1].

For example

Fighting scam ads in Taiwan (2024). In March 2024, AI-generated investment scam ads impersonating Jensen Huang, head of the chip company NVIDIA, flooded Facebook. The Taiwanese government sent 200,000 random text messages, and from those who responded selected 447 people matching the makeup of the population. They discussed the problem at tables of 10. A gathering like this, where randomly selected citizens discuss how to respond to AI and other technology and sum up their views, is called an Alignment Assembly [3].

The resulting proposals fed into legislation, and Tang writes: "By 2025, scam ads were down by more than 94%" [3]. However, the FAQ (frequently asked questions) on civic.ai, the official site of the 6-Pack, narrows this to identity-impersonation scam ads "in the categories those measures touched" [4]. And an editorial note dated September 2026 carefully restates the assembly's role: not the author of the law, but "a response path while a bill was already moving" [5]. Tang cites this case as an example of the first principle, "Attentiveness" [3].

AI in social care in the UK. In Green's work, more than 100 people — people receiving care, carers, care providers and others — jointly created a definition of "responsible AI in care" starting from what matters to them. civic.ai also presents this as an on-the-ground example of the first principle [1][6].

ROOST (2025). A tool for detecting child sexual abuse images, published openly so that anyone can use it. In the manifesto, Tang presented it as a way for online services such as Bluesky, Roblox and Discord to train small AI models suited to their own users' cultures, instead of relying on a single company's detection AI [7]. civic.ai's pages on the fifth and sixth principles also cite it, but add that it is not yet a network in which separate AIs share threat intelligence or train one another [15].

How it works

"Top down" is not enough

Mainstream AI alignment (research on making AI act in line with human values) works from the top down: write better rules, infer better preferences, train better models. The 6-Pack does not reject these. As civic.ai puts it: "Those tools matter. They are not enough on their own" [1].

civic.ai takes the position that "Alignment is not solved once. It is maintained in public" — an accountable civic procedure it calls "alignment-by-process" [1][8].

Turning the ethics of care into machinery

The six names come from the "ethics of care" — a strand of ethics that puts relationships of caring for one another at its foundation. They build on four phases of care set out by the ethicists Berenice Fisher and Joan Tronto in 1990 — caring about, taking care of, care-giving and care-receiving — and on "caring with", which Tronto added later [1][8].

civic.ai says the value of the framework "is not six new virtues but care translated into institutional machinery — and if a pack name does not change who can object, who must answer, what gets logged, or what happens after failure, it is decoration" [1][8].

The six principles

#	Principle	The question it asks
1	Attentiveness	What are the people closest to the problem seeing that institutions still miss?
2	Responsibility	Who is accountable, with what authority, and what happens if they fail?
3	Competence	Does it actually work in practice — audited, explainable and safe to fail?
4	Responsiveness	Can affected people contest outcomes and force repair?
5	Solidarity	Does the ecosystem reward cooperation, exit and public accountability, rather than locking users and data in?
6	Symbiosis	Does the system stay bounded, local and sunset-ready instead of hardening into permanent rule?

Principles 1 to 4 form a loop: "Attentiveness → Responsibility → Competence → Responsiveness → back to Attentiveness". Principle 5 "scales that loop across organisations". Principle 6 is the outer frame — in civic.ai's words, "the boundary condition that keeps every deployment local, plural, and sunset-ready" [1][8].

Each principle also has public measures. For example, principle 2 looks at the share of commitments kept, and principle 6 at whether handover or shutdown is ready. The measures are designed to check whether things are "good enough", not to be maximized, and each setting decides for itself what counts as enough [9].

Small AIs called "Kami"

The unit of AI in the 6-Pack is not one giant AI governing everything. It is a small AI that serves a particular place or group, called a Kami — "a bounded local steward, not a universal governor" [1]. The name borrows the form of a kami in Japanese Shinto: a local deity dwelling in one river or forest. AI governed by the 6-Pack as a whole is called Civic AI.

How the names changed

In the 2025 University of Oxford announcement, the principles had slightly longer names, such as "Attentiveness in Recognition" [2]. The manifesto was delivered as a talk at the AI lab Google DeepMind in London on September 1, 2025 [7][10]. In March 2026 a Civic AI Conference was held in Oxford, and the book Civic AI: 6-Pack of Care is due from Wiley in February 2027 [1].

What is debated

It does not solve the core problem of AI. Readers on the online forum LessWrong pointed out that the 6-Pack merely supplements existing AI safety research and assumes that AI is already roughly aligned with humans. Tang replied that alignment has a "vertical" problem — making a single AI faithful — and a "horizontal" problem — keeping a society of faithful AIs from falling into conflict — and that the 6-Pack addresses the latter [11]. The civic.ai FAQ also states plainly that it does not claim to solve the problem of aligning frontier AI with humanity's values; it is built for a different question: "who, in this particular room, is owed an answer by this particular AI system" [12].
Too soft amid the race for profit? In the same discussion, some asked whether cooperative AI could survive fierce competition between companies. Tang and the FAQ respond by pointing to ways of changing the incentives rather than relying on goodwill. These include interoperability (connecting different companies' services), data portability (being able to take your data to another service) and public procurement (governments using what they buy to favor cooperative AI) [11][13].
Did it only work because it was Taiwan? The FAQ acknowledges: "The mechanisms transfer, but the specifics do not." It also says the framework has yet to be tested in countries with weaker civil society or deeper divisions [4].
Revised claims. The 2025 manifesto now carries an editorial note dated September 2026, correcting or narrowing claims such as the assembly's role and the scope of a certain state law's (a US state's law) effects. Care is needed not to quote the figures and causal claims of the talk version as they stand [5].

Where it fits in Plurality

The 6-Pack came after the book ⿻ Plurality (2024) and does not appear in it. civic.ai itself describes the 6-Pack as "an application of ⿻ Plurality to AI governance" [1][8]. In other words, it is not a claim of the book, but a development from 2025 onward by one of the book's authors, Tang, together with her co-author Green.

There are links with the book. The book already cites Alignment Assemblies as an example of Taiwan's digital democracy [14], and the 6-Pack makes them a central tool. The broad listening (ways of listening to many voices) and bridging (systems for finding common ground between people who disagree) used in the first principle also connect to the discussion in Part 5 of the book [7].

Articles to read alongside:

Civic AI — AI governed by the 6-Pack
Kami — the small AIs that are the unit of deployment
Reverse alignment — the agenda of Glen Weyl and others for "fixing the society side"
Alignment Assemblies — forums where citizens discuss AI

Sources

Civic AI — 6-Pack of Care (home page) — https://civic.ai/ ［CC0］
University of Oxford Institute for Ethics in AI, Accelerator Fellowship Programme announcement — https://afp.oxford-aiethics.ox.ac.uk/article/civic-ai-and-6-pack-care-reimagining-ai-alignment-relational-health
Audrey Tang, "AI Alignment Cannot Be Top-Down" (2025-11-03) — https://civic.ai/ai-alignment-cannot-be-top-down/ ［CC0］
civic.ai FAQ, Q7 — https://civic.ai/faq/#faq-7 ［CC0］
"6-Pack of Care: A Manifesto", editorial note at the end (September 2026) — https://civic.ai/manifesto/ ［CC0］
Caroline Green, "Care AI" — https://civic.ai/care-ai/ ［CC0］
"6-Pack of Care: A Manifesto" — https://civic.ai/manifesto/ ［CC0］
civic.ai Glossary — https://civic.ai/glossary/ ［CC0］
civic.ai, "Map & Measures" — https://civic.ai/measures/ ［CC0］
LessWrong, "⿻ Plurality & 6pack.care" — https://www.lesswrong.com/posts/anoK4akwe8PKjtzkL/plurality-and-6pack-care
Comments on the same post (2025-09-02) — https://www.lesswrong.com/posts/anoK4akwe8PKjtzkL/plurality-and-6pack-care
civic.ai FAQ, Q17 — https://civic.ai/faq/#faq-17 ［CC0］
civic.ai FAQ, Q1 — https://civic.ai/faq/#faq-1 ［CC0］
Plurality, 2-2 "The Life of a Digital Democracy" — https://www.plurality.net/v/chapters/2-2/eng/ ［CC0］
civic.ai, "Pack 5: Solidarity" and "Pack 6: Symbiosis" — https://civic.ai/5/ , https://civic.ai/6/ ［CC0］

We welcome error reports and questions ("What is…?"). Submissions become public issues.

Report an error in this article ・ Ask a question