"Benefits Starting at Age 68, Utilizing 100 Trillion Won in Excess Tax Revenue"... National Pension Reform Plan Co-Developed by AI and Humans

Demonstrating an AI Joint Model Developed by Experts Over Six Months
Expert Opinions Assessed, AI Presents New Alternatives
Conclusion: Raising Benefit Age and Utilizing Excess Tax Revenue

A policy experiment was conducted in which human experts and Artificial Intelligence (AI) collaborated to tackle one of the most challenging issues facing Korean society: national pension reform. Going beyond simple statistical predictions, the AI analyzed real-time votes and subjective textual input, acting as a "Policy Designer" to mediate conflicts and propose a realistic 'third alternative.'

On August 12, the Korean Association for Policy Studies held its summer academic conference at the National Asia Culture Center in Gwangju Metropolitan City. As part of the event, a special session titled "Pension Reform - AI Deliberative Democracy Policy Experiment" took place. Since March, a team of pension experts—comprising Professor Yang Jaejin of Yonsei University, Professor Kim Taeil of Korea University, Professor Min Juhong of Kangwon National University, and Min Hyosang, Research Fellow at Gyeonggi Welfare Foundation—along with an AI expert team that included Professor An Junmo of Korea University, Professor Hong Areum of Kyung Hee University, and Yang Byungseok, CEO of Nextain, developed an independent inference and evaluation model called the "KAPS Policy Decision Model." This session was an experiment applying this model to approximately 250 participants on site.

Real-Time Plan D Derived Under System Overload: “Benefits Starting at Age 68, Utilizing 100 Trillion Won in Excess Tax Revenue”

AI classified the results entered as subjective text by on-site participants in policy experiments into five categories. Photo by Oyu Kyo.

Three major domestic AI engines with proprietary foundation models—LG AI Research's "K Exaone", Upstage's "SOLAR PRO 3", and Naver's "HyperCLOVA X"—were deployed, launching a collaborative and validation system among large language models (LLMs). As 250 participants evaluated three reform proposals via smartphones or laptops, the AI analyzed the responses to generate a new alternative in real time.

The three alternatives presented by the expert team were based on last year’s reform (contribution rate of 13%, income replacement rate of 43%). All required permanent reserve maintenance to prevent fund depletion. Plan A (Stability-Oriented) kept the benefit age at 65, aimed for a target reserve return of 5.5%, and called for an annual government contribution worth 0.6% of GDP. Plan B (Fiscal-Saving) proposed raising the benefit age to 68, injecting an initial 100 trillion won in excess tax revenues, followed by an annual government contribution of 0.25% of GDP. Plan C (Return-Maximizing) would also raise the benefit age to 68 with a target return of 6.0% and no additional government contribution.

The real-time "Plan D," ultimately derived on site after AI-led analysis of subjective opinions and cross-validation by three LLMs, called for maintaining an income replacement rate of 43% and contribution rate of 13%, raising the benefits eligibility age to 68, a target return of 5.5%, injecting 100 trillion won of excess tax revenue, and an annual government contribution of 0.3% of GDP. While similar to the expert team's Plan B, this "pragmatic compromise" more finely adjusted the annual government input to reflect public demands for fiscal stability. Professor Kim Taeil, from the pension expert team, commented, “Plan D derived by AI is a modified version that recalibrates fiscal input ratios according to citizens’ preferences.”

System Overload Causes Confusion: “New Possibilities for Deliberative Democracy Found”

Plan D derived by AI through field opinion collection during the policy experimentation process. Photo by Oyu Kyo.

As this was Korea's first real-time policy experiment, technical difficulties followed. With more than 250 participants accessing simultaneously, there were delays in wireless network (Wi-Fi) performance and server overload, which impeded real-time data processing. As a result, some portions of the originally planned 10-stage participation process could not be completed.

Professor Hong Areum, from the AI expert team, explained, “To prevent hallucination, we implemented a triple-layered guardrail and search-augmented generation (RAG) technologies, ensuring that alternatives were explored strictly within set rules.” She added, “Unlike simulations during pre-verification, the technical challenge of processing large volumes of concurrent traffic in real time at the venue is an issue to be addressed in the future.”

Despite the constraints of system overload, the experiment demonstrated the potential for AI to function as both a "conflict-mitigation mechanism" and a "facilitator of deliberation" by helping reduce intergenerational and income-level conflicts. Lee Seokhwan, President of the Korean Association for Policy Studies, emphasized, “Whereas it previously took over two months for experts to develop policy alternatives, with AI, we were able to compress this process—incorporating input from 250 citizens—into just one hour. AI is not replacing humans; rather, it shows its potential as a facilitator of deliberation, helping orchestrate social consensus and mediate conflict.”

The association plans to supplement the incomplete data analysis caused by traffic issues and officially release the findings on its website. By integrating both offline and online data, a formal research paper on "AI-based Policy Governance" will be presented at the upcoming autumn academic conference.

This content was produced with the assistance of AI translation services.