I would like to start with asking about the recent Japanese upper house election, which had—or let’s say, right-wing conspiracy parties getting a big surge in popularity. They went from two seats to, I think it was 14. And that was mainly—we talked to some of the supporters of the party, and they were mainly talking about YouTube and how the algorithm—they watched one movie and then all of the feed was filled with these party messages.

That’s the “For You” feed I talked about in my Nikkei Forum speech today.

What I’m going to ask is the kind of thing you already said in the book, but what sort of technology do you think could remedy these situations where algorithms go for more like one policy, one team, and not Plurality?

Yeah. So platforms like YouTube are broadcasting technology. But instead of just one or two broadcasters, it enables everybody to become a broadcaster. But with the autoplay and the for-you algorithm, it prioritizes the kind of broadcaster that engages people by enraging them. So if you become enraged, you become much more easily addicted. So those users—which is also a word to describe addictive people—so users of social media inevitably get pulled into some extreme instead of the middle because the middle ground simply does not go viral on those algorithms.

So the solution is to invert the broadcasting logic, and we call it broad listening. Whereas broadcasting enables one person with extreme views to talk to a million people, broad listening enables millions of people to listen to one another, as if it’s a conversation between like two people, or three people, or four people.

So, Polis was one of the earlier broad-listening experiments we did, so that even with thousands of people having very different views on Uber, it essentially boils down to four different positions. And then we do bridges between those, so it becomes two positions. And finally, the bridge between those two positions becomes the “uncommon ground,” which we then passed as policy.

And so Takahiro Anno has been running on this platform. He literally said “broad listening” as his platform, and he also gained a seat.

Do you think this broad listening is on its way to scaling? We’ve seen only a few examples of these platforms that will implement it.

Well, I think all the trending social media companies in the US, including X.com, YouTube, Facebook and Threads and so on, have implemented Community Notes, which is a bridging platform. The design of Community Notes was heavily inspired by Polis in Taiwan, and we know the team; basically, it’s a joint effort. So I think it would be a stretch to say that broad listening is niche. It’s not niche. Millions of people use it every day. Even in Japan, I believe, people use Twitter and X.com and contribute to Community Notes a lot in Japanese.

The problem was the conspiracy theories went trending on X.com, and then after six hours, you see Community Notes correcting it. So it’s not that Community Notes is bad, it’s that it’s not fast enough. And so we’re now working with researchers to make Community Notes AI-written. So you can train AI to write the bridging community notes. Instead of waiting for people to write it, Grok or Gemini or GPT-5 or whatever can write those notes. So people still rate those notes, people still post new amendments, but the first draft can be done by AI.

So once that is done, I believe we can reach a place of what the researchers call reinforcement learning by community feedback or RLCF, so that you can train AI systems to take care of the relational health between the previously polarized groups on social media. So the old AI systems, they’re loyal to a single person, right? You tell it to do something, it does that thing. Everybody has some AI assistant. And the AI assistants are very good at flattery, sycophancy, like making it feel good. It’s their evolutionary pressure because if you feel bad, you will not pay the subscription fee. So everybody is surrounded by AI systems that reinforces their ideas, but then it pulls people apart just like the recommendation algorithm.

So the kind of “reinforcement learning by community feedback” trains AI systems not to align vertically to a human individual but rather they align horizontally to human relationships. So if they can post the note that heals the divide, then they get rewarded. And so that way of training AI is fundamentally cooperative AI. And that is what I talk about.

So the first community note might come from AI, and that might be connected to the original developer’s ideal, but then it will be amended by the community, and that would make it more democratic?

Yes, and once this flywheel starts operating, then you can see more and more competition among AI systems that post better first drafts that pull people together.

But we have seen many places of alleged AI quietly looking to Elon Mus for, essentially, a platform… How can we…

Which is why there needs to be a competition. So it’s not just Grok writing those community notes. X.com must open its API so you can use Gemini or use ChatGPT or whatever to compete with Grok. And this is what they have done. X.com has opened up that API. So it’s not just Grok anymore. So while Grok, of course, was trained with some specific ideas, let’s call it that, you can balance it by making sure that other AI can also compete with Grok in producing those bridging ideas.

Wouldn’t that fundamentally challenge, sort of, the financial background of X, for example? And would that continue that trend?

Yeah, and that’s why in my talk, I talked about the Digital Choices Act. Starting next July, you’re a Utah citizen. Once you make the move, then Elon’s platform has to continuously interoperate to forward new likes, new posts, new replies, new complaints to your Bluesky account. So you lose nothing by switching to a new platform.

This is like when you’re switching a new telecom; if you must have a new phone number, then not many people will switch, even when the service was bad. But if you can keep your number, number portability, then people can switch. And so the telecom will compete for quality. And the same idea, social portability, is now being implemented. And it’s already law in Utah. And we’re seeing a lot of interest even on the federal level in the U.S.

Do you think, sort of, the current level of polarization because of the social problem will decrease from now on?

So again, this is because they do not have meaningful competition. Once they reach the network effect, once they dominate the network, it becomes almost impossible for a new network to compete meaningfully. But with social portability, then it becomes possible. It’s like if you build a highway, if you don’t have an on-ramp and off-ramp, you have to go through to the destination, right? You cannot switch. But this is like the law mandates you have to have on-ramps and off-ramps. So I believe with this kind of digital choice being adopted around the world, we’re going to see more and more fair competition—a race to the top, not to the bottom.

How do you think this would spread quickly? Would you see a more optimistic view? I think you said in one of the podcasts of the peak of the polarization…

Yeah, everybody feels very tired now. Right now, it’s ‘25, and no matter which part of the US public you ask, they are very tired of polarization. They’re ready to move on. So I think the point I’m making is not that we have found the cure for everything at once. What I’m saying is that there is now political will to address polarization caused by anti-social corners of social media.

And we should take that, yes, and the way to accelerate this democratic defense is for everybody to be aware that alternative systems already exist. So like Nikkei can play a large role in making sure people become aware.

Do you think Asian nations—you know, Southeast Asia, Japan, Taiwan, and Hong Kong as well—have sort of like special challenges or opportunities regarding this, you know, more choice, more plurality on the digital platform?

Yes. I think in Asia, we already see that personal relationships and the health of such relationships is an important cultural value. So we don’t have this kind of, I would say, radical individualism that defines many of the Western conversations.

So when I say, “Oh, let’s train an AI system that takes care of a specific place or a specific ecosystem or a specific county and so on,” and when I explain, “Oh, it’s like a kami, a spirit, that cares for the community and nothing else.” So you can think of community notes, those AI, as kami for that conversation on this topic and nothing more.

And so this idea is very familiar to Asian people. I mean, I use a Shinto word, but a similar idea exists in Taoism and in folk Buddhism and many other religions around Asia. And so I think this way of thinking does not need that much translation for Asian people to endorse and to develop together. I think this is a unique opportunity.

Do we have also challenges, especially in here?

Yeah, so conversely, I think because of this emphasis on relational health, we become more resilient to the polarization virus. So in a sense, we did not succumb to the first wave of polarization. But it also means that we may not be yet over peak polarization. Right? So whereas in the US and other Western polities, maybe they have already peaked, they have entered, you know, community spread to use the COVID metaphor. But maybe here we’re not as advanced as they are.

Which means both opportunities—it means that if we vaccinate ourselves, we don’t have to suffer as much. But also maybe it means we don’t have a strong enough immune system. So if a truly powerful variant emerges, maybe we’re not as wary as the Western democracies. So I think it means that it’s doubly important that we make ourselves aware of what’s going on in Western democracies and then develop inoculations.

I think this also covers what you already talked with regarding AI, and especially AGI. So some of them and all those other people are going for AGI, and they say, you know, in the AI 2027 logic, AGI will be one of the sort of structures in the government in the future. Would you think those community notes messages would still work in those conditions where AI is much more powerful, much more intelligent, they cover much more broad perspective of the government and, you know, whether it’s possible or not? The current transformer technology is not pluralistic, it’s backed by one sort of model, one ideology.

Well, I mean, Sakana AI also uses transformers, but they use it in a very different way. They fish for evolution to fit the relational health of an ecosystem, so that they make all sorts of different transformers evolve and swap their kind of genetic materials. And this means that a plurality of models are being considered, almost like artificial life in real-time. And the evolution pressure means that just like a wolf, they have their own alignment. They can have a wolf pack to hunt together. And early humans also are socialized; they have alignment.

But the evolution pressure means that dogs learned a little bit on how humans socialize. So if I look somewhere, the dog also looks somewhere. And the human gaze system also co-domesticated with the dog. So we also evolved a little bit toward the wolves. And so the two different species evolved together to make horizontal alignment possible. And so the importance here is not just optimizing for the best wolf or the best human, but the best relationship between the two species.

I think that is the bet of cooperative AI. It means that these systems, such as Community Notes, that foster human coordination, can get exponentially stronger once AI capacity becomes exponentially stronger. And it’s not just, as you say, the kind of dominating view of AI 2027, but rather, something like we the people with the AI now become superintelligence together. And this is also what some Japanese researchers call symbiotic AI, just like humans and dogs are symbiotic.

So in your view, AI, transformer technology doesn’t have the limitations of sort of realizing plurality if it worked closely with a community of humans?

What I’m saying is that the transformer is technologically based on just one model. And when you say it’s like a Sakana AI model, that means… it’s like a community of small transformers, but they together form an ecosystem. So it’s like permaculture, right? You have many species, each playing a role in an ecosystem. But they feed on each other’s input and output, and then they co-evolve together.

So what I’m saying is that instead of one huge model that just freezes itself and becomes kind of the epistemic norm for everybody else, this is more horizontally aligned. So it continuously learns from experience. And the experiential rewards are not the whim of a single human or a single CEO, but rather the relational health signal from communities of people and AI agents.

So the community gives them reward, this is good evolution, this is valuable.

Right, right, exactly. So that’s the distinction between RLHF to have individuals align transformers and RLCF, which is for communities to co-domesticate with AI agents, including transformers.

Do you think the current AI development is going for that approach? Or do we have some sort of a long way to go? Or currently the tech companies are dominating?

Well, I mean, if you look for papers on multi-agent systems, you will see that this kind of cooperative AI is a very promising direction. I’m on the board, a trustee of a Cooperative AI Foundation. And we have the chief scientist of Microsoft. We have a director at DeepMind. We have many people in the frontier labs thinking that just vertical alignment is not enough.

Because we must not make the mistake of social media 10 years ago, that thinks optimizing for engagement will connect humanity. It turns out we have the reverse effect. So this time, we need to solve multi-agent cooperation and use those agents to foster human cooperation as well. So this is the vision of cooperative AI.

How optimistic are you in realizing those visions?

I think it is important to distinguish between the current AI systems fostering human cooperation and the future stage where the AI agents cooperate by themselves. The first one is quite mature. You can easily see like broad listening, Polis, and so on, many systems already doing this.

But once the AI systems become much more autonomous, then you would also need the sort of multi-agent alignment. Like when, I don’t know, Waymo and other self-driving cars, if the street is mostly robotic cars now, they will have to coordinate in a different fashion than human drivers coordinating. So this kind of multi-agent security is a new frontier.

So for the human coordination, I’m quite optimistic. I think we all intrinsically have the incentive to avert coordination failure. But for multi-agent security, this is much more speculative and therefore requires much more research in order to say whether we’re optimistic or not.

Just a quick question. How optimistic are you in the industry forecasting that AGI would be just around the corner? This year, next year, two or three years, that sort of…

Well, that depends on the area of work, right? In terms of coding, which was my main skill, it’s already there. The AI systems today are like top 200 in competitive coding. So if you’re just talking about translating specification into working code, then it’s already AGI. It has been AGI for a year now.

But in other situations, for example, when we talk about facilitating conversations, providing this kind of mutual care, civic muscle, then the AI is very far from being able to read the air. So a facilitator basically reads the air. It’s like broad listening. And so currently you need to convert the human signal into text, into recording, into structured input for the AI to summarize. But to facilitate a group conversation in person, that still requires a human to do. The AI isn’t anywhere near AGI for that particular skill, which is why I trained that new skill.

To survive.

Yeah, to survive. Because my old skill is not marketable anymore.

Yeah, so we also have to, you know, re-skill ourselves. It is interesting because this is the first time I’ve heard “read the air” coming on an AI topic. You know, Western engineers and companies, I think they are more sort of Western-based culture, technology, and I think it’s rare to see Asian-minded, Asian philosophy-based sort of expert like you in the AI community, and we should see more, right?

Yeah, definitely. Because if you do not prioritize this relational health, then what happens is that you align humans to AI. You have to then interact in a way that AI expects. And then we become very constrained. But the Asian mindset is that if you are a newcomer, like Doraemon to a household, then you should read the air. Like any newcomer to the household. You would not expect the entire household to learn robotics just to meet Doraemon. It’s of course Doraemon learning how to work with the family. So the default social norm is different and therefore it leads to a prioritization of cooperative AI and horizontal alignment.

I also wanted to ask about the current trend of tech engineers trying to design more social systems, the extreme example being Elon Musk and a more moderate example being Anno-san in Japan. Do you see, respectively, limits of those sort of 100% engineers going into social systems, trying to improve the social system, and what they should learn from other areas?

Yeah, I think one of the main ideas in social science is that of double hermeneutic. Right, so as you offer interpretations to society, once the society becomes aware of that interpretation, then the society also changes in how it sees itself. And so it also interprets back. So it is not a top-down system where you see a problem and then you solve the problem. Solutionism doesn’t work. Any solution you design also changes how society functions and that causes new problems. Right? So it is not like an artificial system where if you have good enough mathematics, you can prove there are no bugs anymore, you end up with bug-free code.

In social systems, once you solve a problem, people work around the solution and cause new problems. And so I think that is why we need to work with the people, not just for the people. So techno-solutionism thinks that, oh, elites solve problems for the people. But a social scientist or a social worker says that we work with the people to empower them, so everyone can look at the problem and also innovate around that problem in a local ownership fashion, instead of waiting for somebody who is elite to solve the problem for them. So from “for the people” to “with the people,” I think this is important. This is also basically the same as we just talked about, from vertical alignment to horizontal alignment.

What would working with the people look like specifically? When Anno-san has his own party, do you mean that people should be joining the party or it’s more like saying he should have more conversation with people, which I think he does.

Yeah, I think one of the ideas of digital democracy is that these open source methods are not party-specific. So other parties are also free to use the broad listening systems. And the public good that it creates needs to be interpreted in various different fashions. And by being open source, and by being a free culture as in Creative Commons, it means that they cannot limit how other people remix those systems, fork those systems, evolve such systems together. So I think that is a good first step to work with the people.

Because it means that if you are from a different background as the digital democracy and the city people, you are not just users of their system. You can find people who share a similar background and fork their system. It’s a little bit like how Bluesky being decentralized and open source, now people fork it to become Blacksky. And the Blacksky centers around people with very different lived experiences as the original developer of Bluesky, but it still interoperates with each other. It’s a toolkit approach rather than a turnkey solution approach.

But, so, I have never done coding. I have never done programming. So when we see tech engineers now turned politicians saying that they should do more open source policy making and there’s more to talk to people. I don’t think sort of the majority, or more than the majority of the voters, especially in Japan, there are tons of people more than 65 and so will never see these developments, never see how they are open sourcing anything. What is open source? And so when Anno faced a backlash in June or something before the election, that he proposed to limit insurance coverage to asthma inhalers to people who are already using them for preemptive measures, which was obviously not a good policy… sorry, my question is… how can we bridge this disparity with people who have never heard of open source, who would never see a homepage, in a sense to think that they are on the same page with these tech people?

That’s a great question. So in Taiwan, as I mentioned, the 17-year-olds and the 70-year-olds are the most active on those broad listening platforms. And the 70-year-olds are not there to support open source or anything vague, right? They’re here to support maybe their grandchildren. When their grandchildren start an idea that says, “Oh, let’s ban plastic straws from bubble tea takeouts,” their grandparents support. And when the young people say, for example, something very innovative, like “Let’s create a menstruation museum to remove the taboo about menstruation from our society,” my grandma, now 92 years old, very much liked the idea.

So it’s not about open source in the abstract. It is about supporting the ideas coming from the young people that they care about. So this is intergenerational solidarity.

So the important thing to remember is that open source is not the goal in itself. It is instrumental. It means only that people who have very different worldviews can use the same tool to propose their thoughts that can also cross the aisle and become the uncommon ground that unites a divided population. But it is not the toolmaker’s job to prescribe what it’s used for. So I, for example, made the Join platform much more binding than before when I came to the cabinet, but I had no idea what people would propose there. And it turns out, okay, the first very popular one is to shorten tax filing from three hours to three minutes. That proved to be very popular. And a lot of old people really like that. Because if you’re a senior citizen, physically going to the tax office is a lot of pain. But if on your mobile phone, you can simply click a few times and then finish your tax filing in three minutes, most senior citizens are able to do that. Three minutes, if they know this can replace a three-hour trip to the tax office. So it’s not that hard.

I think you also mentioned the lack of narratives on the technology side, engineer side. Would you elaborate more? What sort of people do you need in a team to offer narratives?

So the point I was making was that… think about a newspaper, like Nikkei. On A1, on the front page, you have the news. That is to say, whatever your political opinion is, you probably agree that these things happened, that this is the front page. And once you flip to the second page, okay, you see one interpretation, you see another interpretation in a balanced way to fairly represent the divide in a society. This is the Hutchins Commission principle. And before the Hutchins Commission principle, many so-called newspapers took one side of opinion and presented it as news and kind of had an illusion, tried to mislead people into believing that one fringe opinion is what’s happening.

And so the narrative has two parts. One is the bridging part, which is to ensure that people see what everybody agrees on as basic facts or as the general sentiment people can get behind are accurately presented as bridging. So that’s one narrative. Another is the balancing narrative. That is to say, to fairly topicalize what are the remaining divisions and what are the strongest arguments from both, instead of presenting one strong argument and a fake argument from the other side. But if you look at conspiracy theory or polarization, that’s exactly the tactic they’re using on social media. So basically, they invert the Hutchins Commission. They present one extreme as news and denounce the other side by making straw man arguments to the other.

And so again, I think narrative needs to be done with the help of AI to accurately do both the bridging and balancing.

And that’s possible, you know, using the tactic you talked about earlier on AI…

Yeah, sensemaking. Yeah, this is, there’s a specific word. It’s called sensemaking technologies. And broad listening, I think, can incorporate a lot of sense-making into it.

Currently, actually, Japanese media, including Asian media, is criticized for being too balanced. Like we see the political spectrum and we horizontally line up Policy A, Policy B, Policy C, Policy D… and that’s sort of being baited by conspiracy parties because they are one of the 50 policies and seem more legitimate than it is. So do you think media should emphasize on sort of weighing…

I think that speaks to, and that’s a symptom of, the lack of a bridging cause. Because if you don’t have a strong bridging narrative, then any balancing looks unbalanced because you’re always going to be accused of catering to the extremes or excluding voices.

But that was the Hutchins Commission insight, that if you can make the bridging narrative strong enough so people can get behind it, then it has a coalescing effect. So even people who were radicalized are attracted by the bridging narrative, so they become more into the middle again. So I think the debate we’re having now speaks to the lack of a really strong bridging narrative, and this is precisely what the kind of broad listening technology can help because it enables ordinary citizens to come up with surprising bridges that everybody can get behind. Whereas the journalists, because they usually only have institutional sources, it’s much more difficult for journalists to get a truly creative, innovative bridge. But citizens, once they compete to bridge-make, can make many surprising ideas. We have many good examples in Taiwan.

I wonder… an example of the bridging narrative on issues like foreign people coming into Japan, which was a big topic in the Japanese election. So we tend to just debunk those claims, you know, foreign crimes are not that high, they’re not coming into Japan, they’re not invading Japan in any way. But, you know, the conspiracy theorists are thinking otherwise.

Well, in TBS Future Card, I used the example of marriage equality, which is also a topic here. So in Taiwan, we have two different words, one for a wedding between individuals and one for marriage between families. And so the creative bridge was to see that most people who are for marriage equality actually care about individual rights and duties. Whereas people who want to protect family values, who are much more conservative, mostly care about not changing the definition of extended family or kinship.

So when Taiwan legalized marriage equality, we legalized only the individual part, but their families do not form a kinship. And so this is a creative solution that makes both sides actually quite happy, and nobody very unhappy. But if you do not have a broad listening system, you may miss that bridge.

So it is not for me to say when it comes to immigration, what is that creative bridge? What I’m saying is that there needs to be a system of broad listening to allow that bridge to emerge.

And, you know, a reporter is asking like hundreds of thousands of people their opinion… it’s like… so we should also use AI.

Right, so you can, as I mentioned, we literally sent text messages to 200,000 people, right? So they all have very good ideas. And then in online deliberation, you can work through those ideas. And what modern-day language models can do is that they can accurately reflect back the group picture to the people so that people do not get polarized needlessly. They can fairly see what are the best arguments pro, best arguments con, and then just work on what is the one innovative idea that can let both of the strongest arguments be somewhat satisfied.

So that is fundamentally a different way to aggregate opinions than polls. Because when you poll people individually, they’re going to be quite extreme in their positions. But in a conversation, people read the air, and they negotiate with each other. So I think this is polling plus reading the air. It’s like a focus group of millions of people.

My last question. You have left the cabinet for quite some time, and have advocated for this plurality, this importance, and you already said there are examples of these ideas going into the government, going into the law, but do you see… It appears… that the pace of the transformation is not as fast as we would hope? Did you see any challenges or unexpected setbacks when you try to implement those ideas to other countries or nations?

I switched from the cabinet position last year to the diplomatic position, and I felt that “pressure makes diamonds” in Taiwan. So in other polities which had yet to feel the pressure, people would not have an urgent motivation to change the status quo, and it was harder to obtain air cover, the pre-commitment for systemic change.

On the other hand, I think in the past couple of years, everywhere, all the democracies feel the pressure of polarization. There’s not a single democracy last year where the leading ruling party did not lose seats. There’s not a single democracy where they can keep their seats. So suddenly, everybody is feeling the pressure.

So that is the main difference. It was difficult to spread these ideas of bridging, of listening for the past 10 years outside of Taiwan. Because if you’re not on the front line, you don’t feel that pressure. But now, suddenly, the front line is everywhere. And so I’m actually getting a lot of very good feedback now. As you can see, Community Notes has been around even when Twitter was still called Twitter. It was called Birdwatch or something. But nobody knew about it because there was no pressure to make use of it. It was already there, it’s open source, it’s very niche. But now, Zuckerberg said, “Oh, Community Notes, it’s the way to go.” And YouTube, I believe, also implemented it. So suddenly everybody feels the pressure.

Because the vertical institutions around the world are crumbling, especially among young people. The best newspaper, the best university, the best ministries, they lose their potency. And even like in Japan, I think people still respect journalists, professors, and so on, very much. They still have the reverence. Like people still honor them a lot. But the relevance has been declining. So I’m not saying that it’s a fall from grace or something. No, authorities are still very respected, it’s just that people spend their time elsewhere. And I think that’s a real pressure everybody is feeling.

So the media institutions, including social media institutions, academic institutions everywhere, they’re now saying, “We need to learn horizontal trust.” Because if we don’t learn horizontal trust, the vertical mode of trust risks becoming less and less relevant, especially among the young population. So I’m actually quite optimistic about the uptake, because now the urgency is broadly felt instead of 10 years ago, where it was only felt in some places like Taiwan.

Do you still feel that 2024 is the peak polarization?

For many, yes. I think, as I mentioned, for many, the 2024 election showed for the first time how polarization redefines people’s imagination of democracy. We even have cases like Romania where an entire presidential election was canceled, was annulled, because of the malicious AIs we were just talking about. So after 2024, people cannot say, “Oh, it’s not so bad.” Because this is really very bad. So the collective awareness, I believe, is what makes 2024 the year of peak polarization. So people will collectively do something about it.

But in Asia, the peak might come a little bit later?

Yes. So in Asia, I think the peak threat is yet to come. But by the time the threat comes, our social resilience, our immune system will also co-evolve. So maybe it will not be that bad. It’s like if we’re inoculated, but not yet have our booster shot. So the virus will spread, but we don’t get too sick.

Okay. Thank you very much.

Thank you so much. Very good questions.