Taiwan Model and AI Agent: Insights from Audrey Tang at WebX2026

At the Limitless stage of WebX2026, Taiwan's Cyber Ambassador and proponent of Plurality, Audrey Tang, participated in a digital fireside chat. The interviewer was Hayate Matsuzaki from the Active Defense Innovation Fund (ADIF). The discussion covered topics intersecting Web3 and digital democracy, from Taiwan's governance philosophy of "trust as soil" to prescriptions for cognitive warfare and safety design in the age of AI agents.

Matsuzaki began by asking about lessons learned from Taiwan's digital governance.

Audrey Tang
In Taiwan, it is said that trust is soil, not oil to be drilled. Over ten years ago, when Uber entered Taiwan, the disputes over ride-sharing were not unique to Taiwan but common issues in many regions. In the vTaiwan project I was involved in, taxi drivers, platform drivers, passengers, cities, and government agencies each held a piece of the truth. Social media only amplified outrage and ignorance like a fog of war, but by widely listening to opinions, we found an agenda together and transformed them into decision-making through predictive methods.
The same operational rules applied during the COVID-19 period, aiding in the tracking of contact tracing apps, understanding mask supply situations, and enhancing resilience for the 2024 elections. Rapidly releasing evidence and allowing civil society to correct it. Not making takedowns, blocking, or lockdowns the default. This is a common formula. It is not about importing Taiwan's answers directly, but rather the process of adapting technology to our own society that is crucial.

Matsuzaki noted that there are no age restrictions in initiatives like vTaiwan and asked about the significance of young generations participating in politics early on.

Audrey Tang
That's right; there are no age restrictions. The only thing to verify is whether there is a local phone number. The DePIN process operates based on contributions, regardless of age or voting rights. What matters when participating is not qualifications but the wisdom brought to the table. Some of the most constructive agenda settings have come from very young people. Those affected by the rules do not always have the right to decide those rules. That’s why broad listening intentionally expands the space and includes them. The role of facilitation is to ensure that those opinions are reflected as part of the consensus rather than discarded as mere noise.
To this end, we have also introduced a reverse mentorship system, allowing each cabinet minister to have at least one advisor under 35 to provide policy advice. When I was a minister, I even had reverse mentors under 18, who later launched Plurality and became my advisors.

The conversation shifted to Japan, with Matsuzaki asking about current initiatives in Japan that Tang is paying attention to, including Takahiro Yasuno and g0v, among other citizen-led initiatives.

Audrey Tang
The main concept I am currently working on is "composability," which means being able to move freely between interoperable different suppliers. There can be a diverse range of models, like schools of fish. There are models that collaborate with generative AI like Sakana, and options like Solana. There is no need to be tied to a specific generative AI supplier. This is not a story of Taiwan teaching and Japan learning; we are also learning ourselves.
I serve as a visiting fellow professor at GLOCOM, where I see mutual learning. Japanese institutions are not just introducing platforms but are willing to teach that technology. When civic tech and gov tech providers participate in local model learning, data is not extracted like oil but handled while retaining sovereignty. Yasuno is a very good example, seriously engaging in this field. I thought Japan was slow in decision-making due to legacy thinking, but in reality, they are responding quickly. You cannot gain trust without giving it. I feel that acceptance is progressing in Japan.

Matsuzaki shared his experience of fleeing from Dubai to Japan due to the war in the Middle East and asked about cognitive warfare, such as deepfakes and misinformation, in the context of Plurality and civil resilience in digital democracy.

Audrey Tang
Taiwan is a region that has demonstrated resilience almost daily against typhoons and earthquakes. We do not believe that each shock can be completely eliminated; rather, we emphasize building resilience itself. I am working with Vitalik Buterin on the concept of "d/acc," which includes protecting democracy. These are not separate things but are under the same banner of resilience.
For example, in Taiwan, the default response to cognitive warfare on social media is not to conduct administrative takedowns. Silencing speech amidst wartime chaos destroys the very trust we are trying to protect. However, when deepfake scams began to appear in Taiwan, we held an Alignment Assembly and sent text messages to about 200,000 phone numbers. Thousands volunteered, and we randomly selected 447 people to reflect the population and held online meetings. This resulted in not only a very successful anti-fraud law but also a new AI basic law that categorized deepfakes and automated opinion manipulation as acts that undermine social trust. I was pleased to hear that a team led by Ken Suzuki in Japan is training similar mechanisms to prevent fraud.
Compared to the very top-down regulations in the UAE that severely punish deepfake and misinformation posters, Taiwan's bottom-up structure has a higher legitimacy. Instead of imposing rules on platform responsibility and KYC from above, we devise rules that we can agree upon ourselves. In Taiwan, people write those rules themselves and accept them as legitimate and appropriate.

Matsuzaki asked about cyber defense in the era of "offensive dominance," where attackers have the upper hand from a preparedness perspective.

Audrey Tang
In the realm of cybersecurity, machines have become so powerful that there are many malicious collaborative agents that can not only discover vulnerabilities but also easily launch long-term attacks that are difficult to detect. They only operate when the "keychain" is complete. Preventing such distributed attacks is very challenging, and the period of offensive dominance, which we wish to be short, continues.
What we are doing is investing in technologies that asymmetrically favor the defense, particularly in the cognitive domain with "pre-banking." We ensure that people can recognize manipulations in advance. In the field of cryptocurrencies, we are investing in "proof by construction." I am currently reviewing the code I wrote in JavaScript and TypeScript 20 years ago, writing invariants and lemmas so that AI can verify them as mathematical proofs. This way, even if attackers have far more resources than I do, the nature does not change.
When delegating to agents, I also impose restrictions that they cannot spend more than X dollars without my signature. This helps curb the lateral spread of cyber attacks like influence operations and jailbreaks. Our system is intentionally designed to be model-independent, assigning three different models—planner, worker, and verifier—to all tasks. Even if one model withdraws from the market, the agents continue to function. Just as removing one Ethereum node does not crash the chain, we can ensure continuity and accountability.

Matsuzaki asked for Tang's thoughts on the idea of renting AI models trained by others, or "borrowing brains," using Tang's own AI agent as an example.

Audrey Tang
We already have autonomous legal entities that can participate in the market and be legally protected like real persons. In a sense, dialogue is already possible with my company, and as an ERC-6551 agent on a permissionless protocol, it has its own Ethereum namespace. By design, this is not impersonation. The g0v community is not trying to become "vAudrey" nor pretending to be my avatar. The scope of delegation is limited, and people can accurately verify the extent of those limitations.
Here, renting a model means providing a service. Typically, learning occurs off-chain, and inference is done on-chain, but we do not extract the model itself. Please compare this to a more exploitative situation where a working repository is automatically uploaded to a central cloud for next-generation model training via a coding harness. Our decentralized approach uses attribution-based control, allowing each data silo to negotiate with the outside world through agents to provide services. This is far superior to a singleton that is updated only once every six months.
For me, the vision of Plurality is to see data as soil rather than oil. Superintelligence does not exist in some data center; it already exists among us. The role of AI systems is simply to enable us to interoperate more smoothly with each other and between communities.

Finally, when Matsuzaki asked for an optimistic message Tang would like to convey in the context of Plurality and Web3, Audrey Tang concluded with the following words:

"We people are the true superintelligence, not tools for destruction. That tool has finally become cheaper than tools for destruction. This is new. Once, software verification was a luxury, but now any community can trust the agents they interact with. Once, deliberation required buildings, but now more than 400 strangers can accurately disagree online. It’s not about increasing agreement but better dissent. Exit, fork, permissionless protocols. Reward those who reconstitute groups after disagreements. The next decade will build this 'merge' by all who inherit it. Keeping data as the soil of the communities that nurture it and maintaining every system in a repairable state. That is how technology ensures no one is left behind."