Their deliberation had real-world effects. Support for a mandatory algorithm disclosure rule fell 27.5 points once participants weighed how easy it would be for fraudsters to game such a policy. The alternative recommendations they offered were built into a parliamentary bill.
The resulting legislative package required digital signatures for advertisers, made platforms jointly liable for scam advertisements, and slowed traffic to the non-compliant. By mid-2026, Taiwan’s Administration for Digital Industries reported that identity-impersonation advertisements had fallen by 98% and investment-scam advertisements by 99%, even as scam losses across the Asia-Pacific tripled.
Taiwan is not a model to cut and paste; but it is a valuable demonstration. It shows that legitimacy and speed are not in conflict: a representative public assembly with a guaranteed response path moved faster than any takedown regime, and left public trust higher, not lower. It also shows that deliberation is a form of civic infrastructure, but one that must be built well ahead of time.
This demonstration has already inspired others. Japan’s Team Mirai party channels always-on citizen deliberation into the drafting of policy in both houses of its parliament. California runs a statewide listening process through Engaged California, which tracks the impact of AI on the workforce. Connecticut has convened a citizens’ assembly on property taxes.
Utah’s Digital Choice Act, in force since July, treats your social graph—your online relationships—the way telecoms treat your phone number. You may take your data to any competing network, and, starting next July, the old platform must honor the user’s request to forward new followers and engagement signals (likes and so forth) to the new one. As the economist Albert O. Hirschman understood, when there is a real possibility of an exit, one’s voice becomes credible.
But if public assemblies and data portability can govern the major digital platforms, someone still must govern the millions of smaller AI systems now entering classrooms, clinics, and councils. For these, the familiar challenge of aligning a powerful optimizing authority with human values in the abstract matters less than a more immediate question: Who, in a given place, is owed an answer when the system acts, and who is authorized to give it?
The answer that my colleagues and I have developed is Civic AI, a framework resting on the ethics of care. The political scientist and philosopher Joan Tronto named four phases of care—attentiveness, responsibility, competence, responsiveness—and later added a fifth, caring with, whose moral register is solidarity. Each step widens the circle: from a person, to a polity.
We propose a sixth, symbiosis, which widens it once more, to the systems that now act among us: a Kami, short for knowledge artifact management intelligence, which is a bounded digital steward whose memory and ledger of actions stay local, rather than being uploaded to the cloud. Every action that the Kami makes must be traceable back to a source/author or be rejected, and every reading session ends with giving something back to the commons it read from.
For example, when verifying a suspicious message, the Kami transparently identifies scams and provides factual alternatives by strictly citing open, auditable sources like citizen databases and official government APIs. It rejects outputting any insights—even if seemingly accurate—that originate from opaque, proprietary datasets lacking a clear public author. This strict traceability guarantees that every piece of information given to the user can be independently verified, upholding the principle of radical transparency.
These conditions do not require any loss of capability. Mozilla’s recent State of Open Source AI report finds that open-weight models now fall within about 3% of frontier performance (and in a local context, they often serve the user even better). The cost of “good enough” has collapsed. When data is tended this way, it becomes soil rather than oil; value grows through tilling, rather than being extracted through drilling.
Still, there is a big asymmetry to address. We are currently building AI capabilities at the scale of national infrastructure, while the capacity to adapt remains at the scale of a pilot project. E. Glen Weyl and Chris White of Microsoft and James Evans of the University of Chicago warn that three undesirable outcomes follow from this trend. We could get productivity without prosperity, execution without verification, and capacity without constraint. Or more concretely, AI is already cutting the cost of watching people far faster than it is cutting the cost of contesting such surveillance.
Weyl, Evans, and White’s “reverse alignment” coalition, which I have joined, focuses on this neglected need. It has already created a list of what we must build and do, none of which requires a treaty.
Case in point, we need personhood and provenance credentials to distinguish humans from synthetic personas without surrendering privacy. Also needed are collective bodies that pay people for the value of their data, on the model of music royalties. We must enshrine data-portability rights in every jurisdiction, as well as maintain standing assemblies with consistent policy-response paths, rather than one-off consultations. Every ministry should have public-participation officers, and public-procurement rules should require open standards, second sourcing, and exit ramps.
At the same time, education must shift more toward rewarding curiosity, collaboration, and civic care, rather than rote output. Taiwan rebuilt its national curriculum around those competencies in 2019, and our lower-secondary students now rank at the top of international civic-knowledge assessments.
When I wrote the preface to the Taiwan edition of Wikipedia founder Jimmy Wales and author Dan Gardner’s The Seven Rules of Trust, I compressed it into a single line: To gain trust, first give trust. The same approach applies to AI policy. Governments and AI labs alike can publish their source weights before being forced to, make corrections before they are demanded, and grant appeals on day one, rather than at the end of a lawsuit.
Democracy, at its best, does exactly that. It is the means by which a society keeps itself correctable. If governed as a mission, AI can widen the scope of policy feedback and keep judgment with the people. After all, we the people are truly the superintelligence.
