AI Replicates Okinawa Governor Election Results with 0.09% Margin of Error | 420,000 Digital Voters Cast Ballots
機械翻訳 / Machine-translated
A
AI Friends
8時間前・14分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What You'll Learn in This Article
A stunning result was announced following the Okinawa gubernatorial election, which was held and decided on September 13, 2026.
Ambiral Inc. ran a simulation of the election results using AI and succeeded in replicating the vote share of winner Genta Koja with a margin of error of just 0.09 points.
In the actual election, Koja won with 59.42% of the vote in his first electoral victory. The AI simulation, meanwhile, produced a result of 59.51%.
A margin of error of 0.09 points represents extremely high accuracy compared to conventional polling methods such as telephone surveys and in-person interviews.
For this simulation, the team generated 420,000 digital voters called "AI personas" that reproduced the voter composition of Okinawa Prefecture, and had each one individually estimate its voting behavior.
The method used by Ambiral consists of three steps.
Step 1: Reconstructing the Electorate from Census Data
First, data from the 2020 national census was used to reproduce the voter composition of Okinawa Prefecture.
Each AI persona was assigned four attributes: "municipality of residence," "gender," "age group," and "occupation."
For example, personas might include "a male company employee in his 40s living in Naha City" or "a self-employed woman in her 60s living in Miyakojima City."
Step 2: Estimating Past Voting History
Next, election data from six previous elections was used to give each persona a "voting history."
This is where the statistical technique known as "ecological inference" was applied.
This method estimates "which type of person voted for which candidate" based on vote-counting results broken down by municipality.
Step 3: Having the LLM Estimate Voting Choices
Finally, each persona's attributes, voting history, and news coverage of the election campaign were fed into a large language model (LLM — an AI that understands text and generates answers much like a human) to estimate "who this persona would vote for."
In other words, the AI determined a voting choice based on information such as: "a company employee in his 40s living in Naha City who voted for progressive candidates in the past three elections and has been following news coverage about the base issue."
Many readers may be hearing the term "ecological inference" for the first time.
It is a statistical technique for estimating individual-level behavior from group-level data.
For example, suppose vote-counting results for a given municipality show "Candidate A received 60%, Candidate B received 40%."
From this data alone, however, it is impossible to know "what percentage of men in their 40s voted for Candidate A."
Ecological inference combines vote-counting results from multiple regions with population compositions to produce estimates such as "men in their 40s most likely voted 70% for Candidate A and 30% for Candidate B."
In this simulation, six elections' worth of gubernatorial data was used to produce these estimates, giving each persona a history of "how they voted in the past."
This allowed the AI to reproduce voting behavior that more closely mirrors reality.
That said, this simulation is not without flaws.
According to Ambiral's announcement, a significant margin of error appeared among voters aged 70 and above.
In the actual election, 35% of voters aged 70 and above voted for Koja, whereas the AI simulation produced a result of 49.5%.
The margin of error reached as much as 14.5 points.
It was also reported that the simulation inaccurately estimated how many people who voted for incumbent Denny Tamaki in the previous election switched their vote this time.
What these challenges highlight is that "AI simulations cannot achieve the same level of accuracy across all age groups and demographics."
In particular, predictive accuracy may decrease in cases where historical data is scarce or where there has been a rapid shift in the political landscape.
The success of this AI simulation has the potential to significantly impact the world of election prediction and public opinion polling.
Challenges Facing Conventional Polling
Conventional telephone and in-person surveys face several major challenges.
First, due to the rise of phone scams, more and more people refuse to answer calls from unknown numbers, causing response rates to drop sharply.
Additionally, the cost per response is high due to operator labor costs and call charges.
Furthermore, because surveys take time to conduct, they struggle to capture rapidly shifting public opinion.
Benefits of AI Personas
AI persona-based prediction, on the other hand, offers the following advantages:
3 Challenges and Risks
However, AI prediction also comes with ethical and technical challenges.
The first is data bias.
If the historical election data or census data contains biases, predictive accuracy suffers. The margin of error seen among older voters in this simulation is thought to be partly attributable to this.
The second is the risk of a self-fulfilling prophecy.
When AI prediction results are made public, there is a possibility that voters will be influenced by those results and change their voting behavior.
For example, if news spreads that "AI predicts a victory for Candidate X," some people may decide "there's no point in voting since the outcome is already decided."
The third is the risk of misuse.
AI makes it easy to create large numbers of fake supporter accounts and manufacture false public sentiment on social media — a technique known as "astroturfing."
There is also the danger that, combined with deepfake videos and fake news, AI could be used to manipulate public opinion.
The evolution of AI technology is poised to dramatically transform the methods used for election prediction and public opinion polling.
The successful simulation of the Okinawa gubernatorial election can be seen as an important step forward in demonstrating that potential.
At the same time, it is important not to rely too heavily on AI and to never lose sight of human judgment and ethical considerations.
How such technology should develop going forward, and under what rules it should be operated, is a question that society as a whole will need to debate.
This article is a cross-post from AI Friends.
