Audrey Tang

You also come to Taiwan every year.

This is a strong convening place. I was just telling Channing how good it is to bring together people who approach AI from very different starting points.

Some people think about AI in almost purely mathematical terms. For them, alignment sounds like a formula that can be solved, after which the system becomes fully reliable. Other people are much closer to practice: policy, competition, institutions.

Three years ago those groups often did not speak the same language. Now, because of SCSP and gatherings like this, more people understand that real AI alignment also depends on the social infrastructure around the system and on public awareness. That awareness is at a very high level today.

But we did have Taiwan Day as part of Day 0.

On Day 0, I especially enjoyed the conversation with Anka Lee (former Deputy Assistant Secretary of Defense) about the civic layer around AI adoption, or the civic muscle that has to be built around it. In the U.K., some people are calling this "muscular adoption."

The idea is straightforward. Instead of assuming there will be one giant AI system in the cloud that everyone subscribes to, people are realizing that this is not always the right direction. A single huge opaque model that folds proteins, folds laundry and claims to fold everything else into one box is not necessarily what communities need.

What we are seeing now is more narrow AI, edge AI and domain-specific AI. These systems let people tune the tool to their own needs.

A good recent example is OpenAI and ROOST's safeguard-model work. The model can compare an AI decision at supersonic speed with a pre-approved rule set and ask whether the autonomous action stays inside or goes outside the boundary.

The important part is that this is small and explainable. It can point to the relevant part of the pre-approved agreement, whether with the Department of War or any other user, and indicate why the action is inside or outside the boundary. It can do that quickly enough to support uses that were mostly speculative a few years ago.

I have been to 28 countries over the last couple of years and I have not had jet lag, only jet boost. But that is not because of agentic AI; it is because of the Timeshifter app, which helps me change sleep patterns with light exposure, coffee and similar cues.

That connects to the earlier point: the useful system is a specific, scoped tool that takes care of a specific relationship.

Taiwan had a large surge in online deepfake scams a couple of years ago. We did not respond by saying that censorship should be imposed top down by the government. Instead, we used what we call Civic AI : focused agents that listen to groups of about 10 people at a time and identify the uncommon ground between them.

For example, Taiwan now has joint liability for social-media companies. If they carry an unsolicited ad that is not digitally signed, and that ad is a deepfake that causes people to lose large sums of money, the platform can be held responsible.

There is also a rule for foreign social-media companies that ignore the liability framework. For each day they ignore it, Taiwan can slow connection to them by one percent.

But the point is not content censorship. The point is that Civic AI helped show broad public agreement around a practical anti-fraud rule. It supported summarization, translation, and social translation, so people from different ideologies could see the same agreement in terms that made sense to them.

The law was made in 2024. In 2025, deepfake impersonation scams in Taiwan fell by more than 94 percent, so that problem is now close to solved.

Because of that, we started exporting the idea. In Japan, not only the LDP but also a new party, Team Mirai , uses broad listening as part of its platform. The people at DD2030 are also training edge AI systems for fraud detection without censorship.

The same broad-listening approach is now in California. Just yesterday, the state launched Engaged California as a listening practice. If you are a California resident, you can take part and share your concerns about how AI automation affects your work.

And the question is not only about jobs. It is also about apprenticeship, belonging, community and dignity. The process can find uncommon ground among Californians and inform legal or economic responses to frontier labs.

I would love to.

The anxiety is justified. Many people's first exposure to strong AI-like recommender systems was on social media.

About 10 years ago, people began seeing those systems push unsolicited information to keep them doomscrolling. The pattern was engagement through enragement. People might post "dear algo, I did not want to see a certain kind of content," but the algorithm did not respond.

The steerability was very low. The system inferred revealed preferences, such as what content kept someone addicted, while ignoring stated preferences, such as the person's wish not to become addicted.

That lack of steerability is a major reason people fear AI.

In Taiwan, broad listening and Civic AI have been used to steer systems toward what people actually want, such as in the Uber case. In places where people do not see that kind of steering, the fear is understandable.

Much of my work over the last couple of years has been with frontier labs on this problem. Anthropic has gathered qualitative input from 81,000 people about their hopes and fears for AI systems. OpenAI and Anthropic have published model specifications or constitutions.

The next question is how to close the loop. If people say they are unhappy with part of a model specification, perhaps because they do not want the model to be so flattering or sycophantic, how does the lab return and show what changed because of that input?

That kind of model-specification or constitution steerability is the frontier. It can turn fear and anxiety into collective agency and civic muscle.

I would point to social translation between communities.

A couple of weeks ago, X rolled out collaborative notes. In the past, if you saw a controversial post on X, you often had to wait hours before a Community Note clarified the context. Now Grok can draft notes that translate between communities.

For example, it might bridge a climate-justice audience and a biblical creation-care audience, writing a note that speaks to both groups rather than only one.

Previously, that mediator role, almost like waging peace, was extremely taxing for a person to perform. Now, the work can be drafted at industrial scale. Research is already showing depolarization when Grok drafts are steerable by people.

This pattern is spreading to Bluesky, and I hope it spreads to Truth Social and other corners of social media. AI-mediated exposure to social media may finally make prosocial media possible.

It is like a campfire. It helps people see one another more clearly.

That is different from the current wildfire dynamic, where everyone is shadowboxing with everyone else.

See you next year.

That is correct. When I was eight — this was 1989 — I started learning programming by reading the books on BASIC and Logo, but without a personal computer. So I started drawing, you know, QWERTY on a piece of paper using a pen, and then I would wake up every day and type CLS , Enter, using an eraser, and then start using a pencil to simulate what a computer will print out.

I think it was an Apple II. And at the time Taiwan is home to many so-called IBM PC compatibles, so my parents got me, I think, an 8088, and then shortly after that an 80286, using those very large floppy disks, which went smaller over time.

It was in fact the same months. So when I was programming on a piece of paper, my father was visiting Beijing for interviews and ended up covering the student demonstrations in ’89, and he went back home on the first of June, which is very fortunate for my family. And he also then, after that, went to Germany to pursue his PhD, to cover the fall of the Berlin Wall, of course, but also his PhD thesis was around autonomous self-organizing networks in the Tiananmen student movement. So I got to know many people who worked on democratization using communication infrastructure, which of course did not fully work in Beijing, as we know, but they did bring that knowledge and it worked slightly better in Europe.

So in 2014 Taiwan was facing a real polarization crisis, because as we know, at the time many people in social media — even though we follow the same people — the algorithm, predicting that we will get more engagement through enragement, starts pushing unsolicited information that amplified the outrage, not the overlap.

And Taiwanese people are already by that time one of the most connected countries in the world, and therefore the most polarized. The president was enjoying 9% approval rating in 2013. So in the country of 23 million, anything he says, 20 million people against him — but of those 20 million people were also very polarized. One side thinks we must be free from China; the other side thinks that we are the free China. And they were really at each other's throats. And so in March 2014 things came to a head when the parliament tried to rush through a trade agreement with Beijing that would open up our telecommunication, publication, cybersecurity sector to investments from Beijing. And of course we think it deserves a deliberation that we didn't get one. So I supported students to occupy the parliament peacefully for three weeks, and we say always that we're not protesting, we're demonstrating — we're showing that it is possible to get half a million people on the street and many more online to actually see each other and agree, instead of being consumed by the antisocial social media like wildfire, to strip away the oxygen and social fabric. We can actually sit around a campfire and have real conversation with people who think differently, in tables of ten. So after three weeks of peaceful occupation, we converged on a set of very coherent demand that the speaker of the parliament simply say, okay, the people's version are better. So let's go with it. So it became one of the very rare occupied movements that converged instead of diverged.

That's exactly right. We essentially just changed one reward function to the social media algorithm. Previously it tried to maximize enragement, because they want to get people who are users — as in drug users — to be addicted to the touch screens. But we instead say no ideas can reach virality unless they speak to the people who are different from the speaker. So only the bridge-makers get virality. And that's really the only thing we changed. And using this kind of pro-social media, people can, despite differences, find their uncommon ground.

Well, indeed, I'm a big proponent of reverse-chronological following feed, as still practiced by Substack and LinkedIn and many more prosocial corners of social media. And with that said, I think pure chronological feed also creates the possibility for people to simply band in their like groups, in their echo chambers. And there's really no way to easily share the cross-group consensus, to discover people who are what we call surprising validators — people who feel differently. Like in 2015, we had a conversation around Uber. Turns out Uber drivers, taxi drivers, their passengers, people in rural places, all actually agree about the core things like insurance and not undercutting meters and so on. But if you're doing purely chronological feed, there's no way for people to accidentally discover that. So I think two layers. One is the normal chronological feed, and another is more like Community Notes, that allows for shared common knowledge around fact-checking and contextualization. That may be the best configuration.

Indeed. I think the best way is to empower people closest to the pain. For example, in 2024, again March, we faced a new wave of deepfake scams on social media, and the incentive was so high for the social media to feature such unsolicited deepfake scams — for example Jensen Huang, the Taiwanese NVIDIA CEO, saying invest in this cryptocurrency. And if you click, Jensen talks to you, but it's not Jensen, it's just a voice clone of him running on an NVIDIA GPU. But Facebook and other social media companies take a much larger share if they run such scam ads. So their algorithms are incentivized to run such ads, because higher profit margin compared to small and medium enterprises.

And I don't really think that they carefully design to prioritize scam. They were just kind of careless in designing their incentives. So we sent—

Yeah, right, exactly. And so the only way to rein them in is to make them lose money if they run scam ads. So we sent 200,000 text messages to random numbers around Taiwan from 111, the trusted official number, saying simply: tell us what should we do about this together. And a thousand people volunteered, and we chose 447 people as mini-public representatives of the Taiwanese public. And then again, table of ten, small campfire, with only one rule: if you think of an idea, if you can convince the other nine people at your table, your idea becomes the national agenda. But if you don't — if you have a radicalized idea — then, well, your idea doesn't bubble up. So people say, for example, if Facebook runs an unsigned advertisement and somebody loses 5 million, then Facebook should pay for that 5 million, because we did not subscribe to it. It's unsolicited. That's a very good idea. And another said, if they ignore our liability rules for every day they ignore us, we should slow down connection to their video by 1%. That's again a very good idea. So we made them into law in just two months, and throughout last year the deepfake ad impersonation ads are down by more than 94%. The problem is resolved in Taiwan. And that is by changing the incentive, not by censoring speech. We never touch the content of the advertisement.

Well, I mean, if it's not financial damage, then it's just reputational damage. And you will need another alignment assembly to discuss that.

Well, I was the youngest digital minister in Taiwan, and also the oldest, because I was the first, in 2016.

Slightly, but not by a lot.