Audrey Tang and Caroline Green in a fireside chat at Protein Studios, 2 June 2026 — introducing the 6-Pack of Care, a reframing of AI alignment that rejects vertical control in favour of horizontal cooperation.
In brief
AI alignment — ensuring AI systems align with human values and serve human needs — remains an unsolved challenge that grows more urgent as frontier labs pursue autonomous and superintelligent systems. In this fireside chat, Audrey Tang and Caroline Green introduce the "6-Pack of Care," a reframing of AI alignment that rejects vertical control in favour of horizontal cooperation. Drawing on established ethical frameworks and proven technology for democratic engagement, the 6-Pack of Care offers a process where AI systems and human communities continuously collaborate to keep human dignity and well-being at the centre of AI deployment throughout society.
Key takeaways
- Alignment by outcome is the default: when a recommendation engine optimises for click-through and profit, it amplifies deepfake scam ads — like Taiwan's fake Jensen Huang crypto pitches — even while "working as designed."
- Alignment by process invites the people affected to steer the system: Taiwan sent 200,000 messages from the trusted 111 number inviting citizens to a jury-duty-style deliberation, rather than banning ads top-down.
- The 6-Pack of Care translates care ethics into design primitives a community can run as a Kami — a bounded local steward, not a universal governor.
- Civic AI keeps human dignity and well-being central through continuous community–AI collaboration, not one-off rules baked into a model.
Watch
Full transcript
Announcer: Here to introduce their bold framework, the 6-Pack of Care, we've got Oxford University's Audrey Tang and Dr Caroline Emmer de Albuquerque Green, give them a round of applause.
Caroline Green: Morning everyone. Welcome to the session on reimagining AI alignment. We're Audrey and Caroline, and Leslie, thank you so much for already defining what AI alignment is all about, which makes my job easier to dive straight into our work, civic AI, when we talk about AI alignment, the narrative often is that it's very much a thing by design, something that is all about the technology and how the technology works. But Audrey, your work has been about AI alignment by process. You very much believe in the civic nature of alignment. Tell us more about that.
Audrey Tang: Certainly. So let's first talk about alignment by outcome, which is the default trajectory. By way of illustration, take a misaligned system. A couple years ago in Taiwan, as well as here, when you scroll on Facebook or YouTube, chances are you see some celebrity trying to give you investment advice or cryptocurrency. In Taiwan's case, it's Jensen Huang, the NVIDIA CEO, Taiwanese. If you click, Jensen actually talks to you. But of course it's not him, it's a deepfake clone of him running on NVIDIA GPUs.
And people lost millions to those fraud scams. But why do those social media algorithms show more scam ads compared to other small and medium business ads? It's because their recommendation engine is aligned to an outcome. Namely, to drive the most click-through and the most profits. Mm-hmm. And so even though it's working as designed, it caused tremendous damage.
And so in Taiwan, what we did is that we built a steering wheel. We did not say, you know, let's stop running ads or anything. We just sent 200,000 text messages from 111, the trusted number from the government, saying, "Let's all steer this together." This is the alignment by process. We invite people who are affected to steer this algorithm toward what people want.
Caroline Green: Okay, so I want to hear a little bit more about that. You sent those text messages to people, and then what happened?
Audrey Tang: So say you got a text message asking, would you like to dedicate, like a jury-duty afternoon's time to talk through this with other people? And you said yes, because of course you have civic muscle, and you will get compensated for your time. So thousands of people volunteered. And then we chose randomly 447 people, a mirror of our population, to have the same demographics as the wider public. Then in tables of ten, each person looking at nine other people on the screen. People started having a conversation.
A civic AI system is like a glorified chess clock with manners. It makes sure that people who are quiet get to speak up. People can disrupt each other, but not for more than five seconds. It shows summary, it shows transcript, and it also measures, for each table, what the uncommon ground is: the rarely discovered common ground that can leave everybody slightly happier and nobody very unhappy in that table. So each table has a civic AI system that reports the consensus of the table. For example, one table says, if we display any ad on social media, let's always label it probably scam, like a cigarette label, until somebody like Jensen can digitally sign and become accountable. Then we take it off. It's a good idea.
Another table said, for an unsigned ad that we did not subscribe, unsolicited, somebody lost $7 million from the ad, well, the social media platform should be liable for the $7 million damage. That's another good idea. Joint liability. The third one: at the time, some foreign media systems do not have Taiwan office. What if they ignore our liability rules? Well, this table said, for every day they ignore us, let's slow down connection to their short video by 1%. Another very good idea.
And so none of this is content censorship, but we were able, at the end of a long afternoon, to put it to vote. And more than 85% of people said, "It's a very good idea." And the other 15% were like, "Okay, we can live with it." Then it became law in just two months. So throughout last year, there are almost no deepfake ads anymore in Taiwan. It's down by more than 94%.
Caroline Green: Wow. So this story, the first time I heard it, I thought that's really amazing. I just really want to go back to the people and how you engage with them. So you send text messages to people, right? Which seems to be something very accessible because, so my background is I work with people in caregiving. So I work a lot with people who live in care homes, who work in home care agencies, family caregivers who, you know, in the middle of the night face something really difficult and often for them, technology is something that can really help, but it's also, you know, an issue of, there are issues of fear and accessibility.
So it seems like a text message is something really easy for people to work with, but tell me a little bit more about that. Is this really a solution to bridge the gaps between those people who are, you know, not normally heard, who are not normally around the table?
Audrey Tang: I think that's a really important point because had we required physically showing up at a town hall for example that would actually rule out a lot of people who don't have the mobility or the extra time for it. But because we made sure that even on a small phone, you can participate in this deliberative polling process, and it doesn't require anything other than an SMS number, and we compensate for your time if you need to buy a headphone or something. So we tried our best to close that gap that you just mentioned, the care gap.
And the other part, of course, is that even if just one person who is relatively shy, but the other nine people are very outspoken, the civic AI system, that's a glorified chess clock with manners, makes sure that nobody stays quiet just because they don't have the experience of speaking up. They would get a direct message — a small nudge — and also make sure that they have dedicated time for their expression.
Caroline Green: Okay, so again, just staying with that example, what struck me too is that Taiwan has very strong civic infrastructure that makes this possible, right? So you were, and here we're talking in our work, which we'll introduce to you a little bit further down the line called civic AI, we're working with a philosopher and political theorist's framework called an ethics of care. Her name is Joan Tronto, who came up with a definition of what care as a practice actually means. In the political sphere and the civic sphere, we have now adopted this framework to help us think through what civic AI really means. And one of the first steps to a good care practice is to be attentive. So here in your example, you know, you were attentive to that need. Yes. Then what's the next step to actually make it work?
Audrey Tang: Yes, we put that into law. So the MPs, of course, have to deliberate. That happens because it's an upgraded poll. We can also show the MPs, all three parties, none of which have a majority in the parliament that this is the very basic thing that people want.
You can add on top of that. But over time, the policy window has really shifted. Because the MPs can really see that regardless of where they live. Whether they're practitioners or whether they're victims of fraud and so on, everybody considered this as proportional. That is attentiveness by having the people who are affected setting the agenda instead of me as the minister at the time or the MPs to set the agenda.
That, I think, is the main difference between this alignment by outcome which is just one abstract number, versus by process, which is a care loop.
Caroline Green: Let's talk about scalability. So, you know, last week we're with a group of people. And we spoke about civic AI too. And someone said, "Oh well, these are great examples from Taiwan, but you know, here in our society, would people really want to engage all the time with important policy issues or so on?" Let's talk about scalability. Is this something that you see we can really roll out?
Audrey Tang: Yes, definitely and in two different levels. One is on an entire city level or municipal level. For example, a couple years ago, there's a young science fiction writer and AI engineer in Japan, Takahiro Anno, who read the book we co-wrote. Plurality — it's public domain, and decided to run for governor in Tokyo. He's just 33, and nobody had heard of him, and no party support. He said, "My platform is going to be broad listening." Using exactly the same open-source technology, he asked everyone in Tokyo to crowdsource his platform.
He is also a VTuber, so he livestreams 24/7 on YouTube as an avatar. You can just call him or a digital doppelganger, but then the real Anno actually becomes accountable to people's crowdsourced platform. Independently, his platform on the eve of election was ranked the best. Of course, he didn't win. He got more than 2% of the vote. But Koike-san won the reelection. Koike-san tapped Anno to run the Tokyo 2050 consultation. That gave him national visibility. Now he is a member of the upper house, a senator in Japan. And in the lower house, the congresspeople include 11 people from his party, the future party Team Mirai. To use this kind of method to build this intergenerational pact. We have already seen it really work on a national level, in the polity even larger than Taiwan in Japan, but also work in family as well.
Just a very quick example. A couple of months ago. My dad had some health issues, and he finds himself talking to ChatGPT much more. And often past midnight, ChatGPT would suggest fantastic cures that are not really scientific just to keep him talking. Because he's a longtime journalist, a very senior editor, he immediately headlined this saying that ChatGPT is only loyal to earn the next subscription from me and not loyal to our family's relational health.
With my parents' consent, my younger brother, Bestian, and I built a small system on a Mac just for our family Signal chat. No other person can reach this, what we call a Kami: Knowledge Artefact Management Intelligence. This Kami really only has one reward function, alignment by process set by mum initially, that says every time my dad talks to this Kami, it should reduce his reliance on the screen and increase his peace of mind so he can find my mum next door being more vivid than a screen. It worked beautifully.
So after a couple months, my dad did get the surgery and is now fully restored to this shared reality. So my point being that alignment by process doesn't just work at civic scale, but also works in your field, in family and home care and aging care.
Caroline Green: Thank you. Absolutely. And I think what strikes me here is that, you know, you're building this as part of your family, but not to replace your family, right? You're part of that, and there's part of care that you cannot replace through an AI, but it can really help you in a powerful way, and in a way that's really aligned with your family values, and, you know, in a space where you know he's also protected, right?
Audrey Tang: Exactly, because it works with no Internet connection, except for Signal or for the peer-to-peer chat. People can really verify that none of our data is going to the cloudy cloud somewhere, but rather it stays entirely within this computer. Also, because it's not running in the cloud and sharing the same processing as thousands of other people, it's predictable. Every time if you enter the same prompt, it's guaranteed to give you the same output. And if you don't like where the output is going, you can steer it.
Tenzin Yangtso here, we cultivated the first Kami together, which also helped us draft the book we're doing. The point here is that because it's so predictable, and it draws only from local knowledge. It has no hallucination problems.
Caroline Green: And I can tell you as an academic, it's a wonderful experience to be writing this book with you and the Kami. The Kami, which is a completely new experience. But yeah, something we can talk about sometime else.
I want to stick with that idea and the concept of the Kami though, and go into more depth because that, the default trajectory at the moment with AI is that we are going to have this overpowering singularity, this, you know, superintelligence that's going to rule our lives. Mm-hmm. And we have very little control over it. Mm-hmm. But what you have described here with these very small, specific systems, that is something very different. Talk us through, you know, some of the key characteristics of that Kami that make it different to the default trajectory?
Audrey Tang: Yes, by this point, there should be a presentation up there, but somehow it's not here. So you have to imagine it, or you can go to the Civic AI website and look at the same image: the illustration, the 6-Pack. Very briefly, what we have learned from the philosopher Joan Tronto is that care is really a loop. We talk a little bit about attentiveness, how to pay attention to people's real needs, and responsibility, how we can build a system so that there's an engagement contract. The whole point of that Kami with my dad is so that he would need it less over time. — that's the responsibility. The competence is the ability to deliver it without relying on lock-in from a cloud vendor, so that's competence. Finally, take responsiveness.
Whenever there's a new need that rises up instead of being locked into the previous trajectory, it keeps all the memories, all the interactions, and so on, but you can steer it directly without waiting for the next version of the cloud model which, by the way, doesn't always work better than the previous version. Rather, you can simply say "Oh, these are the kind of good outcomes we want to see." And then within a minute or so, it just steers itself, called directional steering.
And then add to that, we also need not to be tied to a Mac Mini, we can switch to, I don't know, NVIDIA Spark or something. When we want to switch it, it's not locked into anything. So that's solidarity, because all the different hardware and software vendors cannot just lock people into a vertically integrated solution. Yeah. It needs to be open.
And finally, the idea of symbiosis. So each Kami system does not overreach. It doesn't build dependence and so on. So if there's another group of people, instead of just enlarging the domain of the previous Kami, we just train a new Kami.
Caroline Green: Okay, so I want to stick with that concept of care that's built within the idea of civic AI here, and that whole care loop. Because what we're seeing is that the concept of care is actually being used and adopted by a lot of those very big tech companies. Oh, here we are. We'll go back to this in a minute. Yeah, please.
So the first four are a care loop, as you were saying. Yes, exactly. So you'll be able to find that on our website. That is, you know, the precursor of our book. Everything put into these beautiful designs. But yeah, let's go back to the concept of care, because you know, we're seeing that being, as said, — adopted by some of the big tech companies that we're now creating caring AIs. But there's such a difference between care as a performance — you know, whether the, for example, large language model might sound caring — and care as actual practice. which also takes, you know, a lot of work.
And that's also why Joan Tronto calls it a practice because you actually have to put something into it. And an important part of it is the responsibility factor here. As well as the responsiveness, so that someone actually says, "Well, I've seen there's a need, I will do something about it," and the person that's being cared for, for example, can then say, "Well, actually, you did something here to help me, but it wasn't good care, and you need to change something." So that's what this care loop and the Kami idea is also really trying to capture, right?
Audrey Tang: Yes, I think what it's trying to capture is a different relationship to our personal data, our human interaction. Because previously, in the cloud age, people think of data as oil, right? So we're just the planktons, and our interactions are being extracted somewhere to aggregate into a distillery, refinery, something, and then become plastic, essentially. But what is lost there, exactly as you said, is that if people are like, "Oh, it shouldn't work this way," sometimes it's already too late. We become like a human in the loop of AI, like a hamster in a hamster wheel. [Laughter] We have to keep running in order to keep up but there's no steering of the hamster wheel. I think what we're doing here in Oxford: we call it data as soil. Not as oil. Instead of being extracted upward, we tend the garden, till the garden, regenerate the data. Because the data is never extracted out anywhere else. It is the Kami that safeguards the relational health. If we do need to transact with the outside world to improve. For example, prediction for health matters and so on. The Kami negotiates the privacy boundaries. So it's not for each individual to get doxxed or anything like that. So I think this is a much better arrangement than the extractive relationship because people can put AI in the loop of existing communities instead of human in the loop of AI.
Caroline Green: So it's a beautiful idea. We also know it works. The technology is there, the examples are here, but it doesn't seem like the world works in line with civic AI. And often when we talk to groups, we get that pushback, right? They'll say, well, you know, how are people going to make money out of that? How is that going to become more powerful than the big tech companies and what they're pushing, putting out there? Where do you see the future going? How can we have more civic AI?
Audrey Tang: Well, I think one big part of this is just the awareness, which is why we are here in South by Southwest. The fact that such ideas exist, and you can easily practice them yourself. For example, even if you're still using ChatGPT, you can invite your friends into a group chat in ChatGPT where it behaves completely differently. It becomes aware of the group dynamic and restores the group dynamic, the health, instead of becoming sycophantic.
Or, if you're using Claude or Gemini or something. You can change the one line of your system prompt, the personality, and say, "Please present me multiple perspectives from my communities in an interactive webpage," and immediately it loses this synthetic intimacy. It doesn't try to be your best friend. You get back a brochure that you can really share with your neighbor, your real human neighbor. Instead of this overly personal conversation transcript that we would be, you know, too ashamed to share.
So there are some very simple tweaks that people can really do. All my screens are grayscale, so you can use a color filter to keep only 20% color. That means the reality around us becomes more vivid than the screens, so we can all sleep better and scroll less. With that peace of mind, then we can entertain better interaction patterns with machines, essentially.
You're also working on a social venture from Oxford, so would you like to talk a little bit about how that idea, this care loop, applies to the field of work?
Caroline Green: Yeah, absolutely. Yeah, thanks for mentioning that. So I'm working on something called Dedicate, which is a platform to support family caregivers in their everyday lives to navigate all these major life changes, the practical issues that they're experiencing. And so we are working with family caregivers, the same way, you know, that you have just described, really trying to understand what their pain points actually are, and then seeing how tech can work for them. However, we feel that what's most important to people is actually the collective around them. It's their interdependence, that there are actually organizations, individuals out there who can help them, but the problem is to find these people and to find the answers and the support in the times when they need it the most, and they're already stretched. So that's our starting point, is the collective. Is the people around them who can help. And then we use tech, and we try that out, and AI, to actually support people to meet each other, rather than trying to replace, you know, the human support.
Audrey Tang: Right, so it becomes just a connective tissue to "we the people", who are already the superintelligence.
Caroline Green: Absolutely. And you know, the problem is also people are very fearful of AI and tech. They don't trust tech companies. They feel like, you know, they're just going to be extracted from. They won't be able to get out of contracts. They don't know where their data is going. And also often — I mean, including me — I find it really tricky, and you saw that last week, Audrey. I was trying to build my own Kami for myself. I was like, how do I do this? I still don't. I don't speak that computer language that you still need. Right? It really needs to be an appliance.
Audrey Tang: Yeah, absolutely. So I think, you know, there's still a lot of work that needs to be done in order to really democratize technology so that people can actually use it without fear, without having to speak that, you know, really be like computer scientists in order to build something like that for them. So it's a big and really exciting journey. So that's for sure.
Caroline Green: But yeah, so coming back to our project, I want to talk a little bit about a trip that we did together with Tenzin earlier this year where we went to Dharamsala to spend some time with the Tibetan Buddhist community to learn from them. And what are you taking away from that trip? What did you find the most interesting, sort of, you know, working with that community?
Audrey Tang: Yeah, so the Dalai Lama endorsed our previous book, the plurality.net book. And in the book, I think it begins with a quote from him. And I quote, "It is under the greatest adversity that there exists the greatest potential for doing good, both for oneself and others.”
And indeed, the Tibetan Buddhist community in Dharamsala is facing a lot of adversity. Not only is their language and their culture being systematically erased where they came from. But also, if you talk to AI models, it just doesn't speak the Tibetan language, the culture, the scriptures, and so on. So they see this enormous epistemic gap — the gap of the knowledge they have locally and also the world's desire to look at technology in a way that Buddhists have already comprehended, but there is an enormous gap between the two.
What they're doing is that they're training these very small systems like a Kami called Monlam AI. It's basically taking only the scriptures, the Dalai Lama's teachings, and so on into a very small model. It doesn't try to fold protein, fold laundry, do everything, Studio Ghibli images, or anything. It just does one thing, which is to take whatever the modern science has to offer, but then interpret it in the Tibetan Buddhist culture. It works very well. It's almost perfect actually when I look at it. And far less hallucination, far more grounded than anything the largest language model can use. It also just runs on something very small, even on a phone. That's the insight I got, which is if we know what we're doing, then we don't need this thousand times more energy. We don't need more — I don't know, power plants, data centers, or whatever, if we know what we're doing so that we can train such small systems. And what did you learn from the trip?
Caroline Green: So we actually got to spend a lot of time in spaces that didn't have anything, or didn't seem to have anything to do with technology at all like a nunnery, for example. A school that was built for the local children from some of the poorest communities there.
And what I learned is, you know, so often working in AI ethics, we've now got that question of, what makes us human in the age of AI. And for me, this is a question. I'm like, well, just look at these spaces, you know, talk to the people. Especially the school that was built by a Tibetan monk who saw the poverty around him and wanted to do something about it and support these children. So that was really like compassion manifested.
And, you know, the way that people are able to change, choices, work around difficult situations. The resilience that we as humans have. And I think that one vulnerability, but also one superpower that we have as humans is hope. You know, even in the worst of all times, in the darkest of all times, we find ways to see a light and to work towards that. And I think that's a superpower to lean into radically at these times.
Thank you. What gives you hope, Audrey?
Audrey Tang: I think the ability for "we the people" to already see ourselves as superintelligence. That really gives me hope. More than 10 years ago, another Oxford philosopher, Nick Bostrom, wrote about superintelligence, in which he outlined all the different ways the takeoff can go wrong. The field has been focusing on how to have a safe takeoff. I think what gives me hope is that with people in Tibet, in Oxford, many places. We're working on a theory of a safe landing. Even if we make the takeoff go well, where do we land? I think we should land in our communities, as Pope Leo XIV said, I think last week in the encyclical, instead of building a tower of Babel, that proposes to solve all the problems for everyone by monitoring everyone. We should just rebuild the wall together, and each community. each family, just focus on the segment very close to them. Then, instead of solving anything for a universal scope, just solve things within the community. And then let AI be the connective tissue between communities.
Caroline Green: That takes me to the last minute, and I think it would be good to do a call for action. for everyone who's here. And I would say that, you know, we've already had people from around the world who have contacted us with projects that they feel align with civic AI and they wanted, you know, to get some input from us or some support or so on. And I would encourage everyone to keep on doing that, to reach out to us and to share any kind of ideas or projects that you have where you feel, "I feel like this is aligning" with your work. I thought that was really beautiful. But what's your call for action to everyone?
Audrey Tang: And if you're interested in what Takahiro or Engaged California — the future of AI in the workplace — is doing in California, there's also an equivalent project just starting here in the UK. And that's nationalstrategy.uk.
Caroline Green: Great. Well, thanks to everybody for joining us, and I hope you have a wonderful day. Thank you, Audrey.
Audrey Tang: Thank you. Live long and prosper.