Generative artificial intelligence (AI) is drawing attention for being used more than ever ahead of the ninth nationwide local elections on June 3.

AI is being used not only for opinion polling but also for broadcasters’ real-time vote count analysis and explanations. AI is also being used for deepfake crackdowns.

◆OpenAI... GPT-5.5 reads the race

AI will be introduced to terrestrial election broadcasts. SBS and OpenAI Korea signed a business cooperation agreement on May 7 at SBS headquarters in Mok-dong, Seoul. They decided to deploy generative AI in real time for the June 3 local election special vote count broadcast, "The People's Choice 2026."

A key difference from past AI vote count broadcasts is who provides the commentary. SBS previously presented AI election broadcasts such as "AI Yoohwakdang" and "AI Tupyoro." In that format, an in-house statistical model produced win-probability figures, which anchors and experts then interpreted and delivered.

This election will introduce an "AI situation room." OpenAI's latest model, GPT-5.5, will receive vote count data in real time and generate natural-language commentary directly. It can also automatically detect closely contested areas and automatically extract timing for when a winner’s outline emerges. AI will also handle real-time comparisons with vote count flows at the same time in past local elections. Nam Seung-mo (남승모), head of SBS' election broadcast planning team, said, "With generative AI-based technology, we overcame the limitations of past vote count broadcasts that stopped at delivering numbers." He added, "It will resolve viewers' questions in real time, such as which areas are close races and when a winner’s outline emerges."

It will also introduce a pledge analysis service called the "AI election secretary." The service is designed to help voters easily understand candidates’ pledges and was prepared as part of the National Election Commission’s policy election promotion project. AI will also create broadcast visuals. Videos produced on a ChatGPT basis by AI artist Choi Se-hoon (최세훈) of OpenAI Creative Lab Seoul will be aired on live TV and YouTube streams. They include a main title and an exit poll countdown video that visualize the broadcast slogan, "Fill the gaps, fill my life."

An OpenAI Korea official said, "This is a good example showing that AI can be used to enhance accessibility to democracy." The official said AI makes it easier to access candidates’ pledges in chatbot form and handles labor-intensive analyses such as comparing past election data. The official added that AI’s role in the collaboration is to make difficult things easy and provide citizens with richer information.

◆Reading public sentiment with synthetic samples... spotlight on 'silicon sampling'

AI use is also spreading in opinion polling. The "silicon sampling" method, which creates virtual survey groups using generative AI, is also drawing attention. If demographic conditions such as age, region, education and ideological tendency are entered into a large language model (LLM), the model infers how voters with that profile would respond based on patterns learned from training data. It combines this with news, social media and search data to simulate political tendencies and issue responses of specific groups. It is being discussed as an alternative that complements low response rates and sample representativeness problems in phone surveys.

In an actual validation study, AI reproduced real voting results with a low margin of error. A polling firm, STI, built a generative AI model that corresponded one-to-one with 2,139 real panelists based on 2025 presidential election data. In a comparison, support rates for major candidates such as Lee Jae-myung (55.4 percent versus an actual 57.3 percent), Kim Moon-soo (31.7 percent versus an actual 33.7 percent) and Lee Jun-seok (5.4 percent versus an actual 6.5 percent) recorded errors of less than 3 percentage points.

Lee Jun-ho (이준호), STI's CEO, said, "While we can match the vote shares of major candidates relatively similarly, it is still not easy to precisely reproduce the actual composition of support bases."

AI services predicting the June 3 local election landscape have also emerged. An AI election analysis platform called "Huwin," created by Kim Jung-tae (김정태), a professor at Dongyang University, was implemented by training on local election, parliamentary election and presidential election data since 2012, along with opinion polls. In Huwin's analysis of the Gyeonggi governor's race, Democratic Party candidate Choo Mi-ae stood at 58.0 percent and People Power Party candidate Yang Hyang-ja at 36.0 percent, putting the gap at 22.0 percentage points. It saw the gap about 8 percentage points wider than the previous official opinion poll gap of 14.1 percentage points. A simulation of basic local government head races across Gyeonggi's 31 cities and counties was classified as 23 areas leaning Democratic Party, 3 leaning People Power Party and 5 toss-ups.

Kim said he used neutral AI technology and designed it to avoid distorting information as much as possible. He explained that it was intended to help read objective, data-based trends.

Under the current Public Official Election Act, however, it is difficult to officially announce AI synthetic surveys. Some concerns also exist that results not answered by actual voters could be disguised as public opinion and affect elections.

◆Deepfakes nearing 10,000 cases... AI makes them and AI catches them

AI has also been deployed in crackdowns on deepfakes that are surging in election campaigns. Deepfake posts for which the National Election Commission asked platform operators to delete stood at 9,956 cases as of May 25, more than 25 times the level in the 22nd parliamentary election, when there were 388. It is nearing the level in the 2025 presidential election, when there were 10,510. The surge is continuing with each election.

A legal basis is also in place. Under Article 82-8 of the Public Official Election Act, producing, editing or distributing election campaigning that uses deepfakes is fully banned from March 5, 90 days before election day, through election day. Violations are punishable by up to 7 years in prison or a fine of 10 million to 50 million won.

A system in which AI catches fakes made by AI has also been put into operation. The Ministry of the Interior and Safety and the National Forensic Service deployed an "AI deepfake detection analysis model" for this election in March. The core is combining the top 5 models from a competition in which 268 teams, or 1,077 people, participated in December last year using an ensemble technique. Detection accuracy for a single model rose to 92 percent from 76 percent, and increases to 97 percent when the ensemble is applied. While existing detection technology focused on facial areas, this model combines global analysis of the entire video with local analysis of specific parts.

The election commission has formed a special response team of about 440 people to focus on monitoring social media and portals and requesting deletions. The National Police Agency's National Investigation Headquarters also raised its election crime investigation system to the highest level from May 14. Yoon Ho-jung (윤호중), minister of the interior and safety, said at a demonstration of the AI deepfake detection analysis model in March, "As AI technology advances, new threats such as deepfakes are also increasing." He stressed, "We will work closely with relevant agencies such as the election commission, the police and the National Forensic Service to actively respond to false and manipulated information and create a fair election environment that the public can trust."