Two rays of light intersect at a massive stone door. Before it stands a small figure. The same door holds two kinds of light—the warmth of consensus and the chill of surveillance. We stand at the place where we cannot yet choose which light to enter.
1. Four Years Ago, and the Question That Has Shifted
In spring 2022, I wrote a short column in a local newspaper as a member of the Yeongyang Electoral Management Commission. The title was "Political Participation for the 2030 Generation Is No Longer Optional." I introduced civic hacking in Taiwan, platforms like vTaiwan and pol.is, and argued that Korean youth must arm themselves with technology and engage in politics.
I still remember a quote I cited from James Surowiecki, a financial columnist. "For collective intelligence to work, you need participation from diverse populations and a system that can block human factors like mob psychology." I wrote that AI and big data would create that 'system.'
Four years have passed. The world has moved much further. What was imagination in 2022 is now a tech company's product roadmap. Google DeepMind published the "Habermas Machine" in Science in October 2024—an experiment showing that LLMs can achieve broader consensus than human mediators. vTaiwan's cumulative participants exceeded 200,000. The UK Collective Intelligence Project's "Alignment Assemblies" expanded to 68 countries.
So my four-year-old question must be updated. Saying "the 2030s must participate in politics" is no longer sufficient. The real question now is this:
At the moment when AI can, for the first time, technically fulfill Plato's philosopher-king conditions—perfect knowledge, unbiased judgment, deliberation without spatial or temporal constraint—are we opening the door to perfected democracy, or to perfected dictatorship?
2. The Ancient Dream Opens
In the 4th century BC, Plato wrote in The Republic that philosopher-kings should govern. Three conditions: possess all knowledge, be swayed by no self-interest, and have endless time for deliberation. For 2,400 years this remained an ideal only. Humans cannot know everything. Bias cannot be engineered away. Time is finite.
So humanity chose the second-best option: representative democracy. Abandon the philosopher-king. Instead, voters elect representatives who, with limited time and limited knowledge, render judgment. Not perfect, but workable compromise.
Then AI arrived. The Habermas Machine, in an experiment with over 5,000 UK adults on deeply divisive topics, produced consensus statements that participants rated as clearer and less biased than those from human mediators. vTaiwan processed complex policy questions—Uber regulations, remote medical care, online alcohol sales—through citizen deliberation and the Taiwanese government ratified those conclusions as actual policy in May 2016. The UK, US, and Taiwan governments, alongside OpenAI, Anthropic, and the UK AI Safety Institute, now operate "Alignment Assemblies" where citizens can directly weigh in on AI policy in 68 countries.
At this moment, Plato's three conditions are being partially met. Knowledge is being compressed by large language models. Bias is being offset through multi-algorithm approaches. Space and time are being erased by networks. The philosopher-king emerges not as one wise person but as a hybrid of citizen collectives and AI mediators.
3. But the Same Door Opens to Perfect Dictatorship
Do not stop there.
In March 2025, China's State Council and Central Committee released a 23-item policy document redefining the social credit system. Hundreds of millions of surveillance cameras, AI companies like Hikvision and Sensetime, facial recognition, voice recognition, and predictive analytics—all integrated into police, transport, medical, and financial systems. This system is spreading through bilateral arrangement to 63 countries. Biometric infrastructure installed by Chinese AI companies in Zimbabwe sends that data back to China, turning Zimbabwean citizens into involuntary 'training data.'
Here sits an uncomfortable fact we must remember. Hikvision's AI and the Habermas Machine use the same transformer architecture. Only the training data and the objective function differ. One is designed to 'score' citizens. The other is designed to convert opinions into consensus.
What then is the branching point between the two paths? Not algorithmic quality. Not data scale. It is the question of to whom power must account itself. Democratic AI must explain to citizens. Authoritarian AI need not explain.
There is something more to watch carefully. Even democratic AI, if citizens cannot trace how their opinion was processed and aggregated, preserves the form of accountability while losing its substance. We are erasing the boundary between genuine explanation and explanatory theater. Multiple South Korean municipal governments are deploying AI to auto-classify citizen complaints. But it remains unclear how transparent those classification criteria become to residents. In that condition, we are building black boxes wearing democratic faces.
Small figures raise their hands and write toward a large question mark carved into a stone wall.
4. Three Rebuttals a Critic Gave Me
I will be honest. Before writing this column, I passed these three sections to a dedicated AI critic with this request: "Rebut me as forcefully as you can, but don't simply negate—make it more complicated." That critic received only my hypotheses and arguments and my own acknowledged weak points—not the full column.
The rebuttals that came back were sharp. Three in particular stung.
First, distortion of the time axis. vTaiwan became an institutionalized policy tool not because the technology matured, but because a political crisis—the 2014 Sunflower Movement—came first. The crisis activated government, opposition, and civil society simultaneously. Technology happened to be ready. Technology did not lead democratic innovation; democratic crisis drew the tool from the shelf. How is Korea experiencing that crisis now? Is it a problem technology adoption can sidestep?
Second, the illusion of scale. No matter how many people accumulate on AI deliberation platforms, they remain minuscule against the millions who flood online petitions and social media opinion. People flow to 'channels with impact,' not to 'structured deliberation channels.' vTaiwan's 200,000 cumulative participants represent less than 1% of Taiwan's population. Calling this a "completion of democracy" may be the oldest trap: using formal democracy to obscure substantive representation.
Third, the most painful paradox. Suppose AI-mediated deliberation succeeds and actually produces good policy rapidly. What message does that success send to citizens? "Policy is made by technology and experts." Voter turnout might rebound briefly, but delegation psychology deepens. My original 2022 goal—strengthening youth political participation—could collapse precisely because AI succeeds.
I cannot fully answer these three rebuttals. If I did, I would be lying. But those rebuttals themselves made clear: democracy does not emerge from technology alone.
5. What Korea's 2030 Real Numbers Are Saying
So I return to the figures in my four-year-old column.
In the 2022 nationwide local elections, the 2030 voting rate was 53%. In April 2024's general election, the 20-something early voting rate dropped to 12.9%, and 30-somethings to 11.3%. In the same period, early voting rates for 50s and 60s exceeded 22%. Youth turnout decline is not a trend. It is an acceleration.
And here is the more decisive figure. In Gallup Korea polling, the proportion of 2030s responding "no party affiliation" reached 54%—an all-time high. Trust in political parties generally has visibly fallen in recent years.
These numbers do not speak to 2030s' political apathy. They speak to the failure of existing parties. Youth have not abandoned politics. They have judged that there is nothing worth choosing. In this context, what changes when we open an AI-mediated deliberation platform? It offers more options to choose from. But the institutions that would receive and legislate on its results—parties and parliament—have already lost trust.
I watched this reality for years as an election commissioner. Voting rates are a mirror of party strength. When the electoral commission intensifies outreach and urges participation, but parties offer nothing of value, voters do not show. AI-mediated deliberation platforms operate the same way. They become democratic tools only atop democratic institutions. Otherwise, an elegant tool sits alone writing consensus to an empty room.
6. What Comes First in Yeongyang: Three Sequences
This brings me back to my column from four years ago—and reveals what needs revision. "Give the 2030s AI tools" is insufficient without institutional design that allows those tools to touch actual power. For a small region like Yeongyang, this sequence is realistic.
First, transparency of the assembly before any external platform. Before bringing in outside AI deliberation, the Yeongyang County Council itself must become transparent. AI should summarize and compare council members' legislative activity so anyone can search what each said, how they voted, what policies they proposed. Internal decision-making must come before external platforms arrive.
Second, regional language data before external platform dependency. The Habermas Machine, vTaiwan—both reflect the cultural context of their training data. If Yeongyang's agricultural products, Yeongyang's complaint patterns, Yeongyang's dialect and lived sense are not well-translated by AI, we end up with Seoul or Silicon Valley answers to Yeongyang questions. The most fundamental autonomy a small county can claim is organizing its own regional language data. Document and prepare data on our place.
Third, a citizens' board for AI decision monitoring at the ordinance level. Every AI decision system the county deploys—complaint classification, administrative chatbots, budget allocation assistants—should have its algorithms, training data sources, and failure cases publicly reported quarterly, reviewed by citizen representatives through a formal municipal ordinance. What is difficult at the national level is possible at the 15,000-person scale. If the difference between China's social credit system and our democratic AI is the direction of accountability, then the institution that demands accountability must exist locally.
A small figure holds up a large key in the place where warm light pours from a half-open stone door.
7. Why We Must Still Trust the 2030s: Updating a Four-Year-Old Conclusion
Four years ago I wrote: "The 2030 generation is far more competent at implementing digital democracy than we assume. Why not give them real work?"
I hold that conclusion, but I must change one thing. Entrusting them does not mean giving them better AI-using ability. It means developing their capacity to demand explanation of what AI produces. Not prompt engineers. Watchdog citizens. That role is cultivated through training, not authority.
So I want to ask three things of each role in our region.
If you are a civil servant, prepare the algorithm explanation document for the next AI system your county adopts in language citizens can read. Not technical terms. Just "by this standard, this decision is made." One page begins a culture of accountability.
If you are an educator, teach students not how to use AI but how to question it. Not "Is this answer correct?" but "On what basis does this answer exist?" That is the basic literacy citizens need in the philosopher-king era.
If you are a young citizen, do not stop at voting in the next election. Choose one county decision process and follow it to the end. Read public meeting minutes. Request withheld documents. One such experience shatters the feeling that "politics is not about me."
Four years ago I was optimistic. Now I hold both optimism and caution. Plato's door is already opening. Whether it leads to democracy or dictatorship depends on what those standing before it have prepared.
Yeongyang is small. But on this question, Taiwan and China and Seoul stand on the same starting line.
Images in this column were created using AI image generation (Google Gemini 3 Pro Image) and underwent editorial review by the Yeongyang AI Association.
This article was produced with the deep-thinking engine
Simthink Engine (深 Think) of Myeongseong Simjae.