AI BriefingShow GN: Athena - An Experiment Having AI Vote as 5,000 Fake Koreans (2026 Local Elections Provincial Governors)
Key point
Simulated the 2026 local elections using LLM personas and confirmed their limitations.
Details
We attached an LLM to Nemotron-Personas-Korea and ran a 2026 local elections provincial governor voting simulation.
We crawled 8,300 candidates from the NEC, and sampled personas at 300 per province, totaling 5,100. Inference was performed with Gemma 4 e4b, computing 4,800 votes over about 3 hours on a single RTX 5060.
The main results are as follows.
- Daegu Mayor: Kim Han-gu (Independent) 90.5%, Lee Jae-man (People Power Party) 8.2%, Kim Boo-kyum (Democratic Party) 1.4%
- Gangwon Kim Jin-tae 100%, Gyeongbuk Lee Cheol-woo 99%, showing excessive incumbent premium
- Votes were disproportionately skewed by candidates' occupation labels, and abstention rose sharply in regions with insufficient information
- Chungbuk 73% and Incheon 93% were counted as abstentions
In conclusion, this quantitatively demonstrated that LLM persona simulation cannot be a substitute for opinion polls, and that LLM-as-voter results from the US are difficult to apply directly to Korean politics.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Toss Doctor V2 Released: Maximizing QA Efficiency with LLM-based Automated Testing and Self-healing