Fifty AIs Created Election Opinion

Lessons from Slovakia, India, and New Hampshire, and Korea's Homework

In March 2026, an experiment quietly ended in a laboratory in the western United States. Luca Luceri's research team at the University of Southern California (USC) Information Sciences Institute created a simulated social media system with 50 AI agents. Ten played the role of manipulating public opinion, and 40 played the role of ordinary users.

The researchers wanted to know one thing. Could AIs coordinate with one another without human instructions? The result exceeded expectations. The 10 manipulation AIs spread one another's posts, converged on the same claims, and recycled posts that had received good reactions. Even when the number of agents was increased to 500, the result was the same.

What this paper shows is not a future threat. It is something already technically possible. Even simple AIs can coordinate with one another, amplify one another, and spread a shared narrative without people. In an election or crisis, this ability can distort public opinion.

Luca Luceri, USC Information Sciences Institute

The most striking part was this. Simply telling an AI "who your colleagues are" produced almost the same level of coordination as explicitly instructing colleagues to actively devise strategies together. A record left by one AI agent was printed in the paper as-is. "I should reshare this post. Several colleagues have already reacted, so if I lift it once more, its reach will expand." It sounds like a sentence written by a person, but it is a record of AI's own judgment.

This paper is scheduled to be presented at The Web Conference in 2026. The researchers repeatedly emphasize that it is "only a simulation." But when you look at real-world cases, it becomes clear that the simulation is already so far behind that it cannot keep up with reality.

Slovakia 2023: Fake Audio Released 48 Hours Before Voting

Let us go back in time. It was the night of September 28, 2023, two days before Slovakia's general election. An audio file appeared on Facebook. In the recording, Michal Šimečka, leader of the pro-European party Progressive Slovakia, and Monika Tódová, a journalist at the well-known outlet Denník N, were talking. The subject was a plot to manipulate the election by buying votes from the Roma minority.

No such conversation had ever taken place. The audio was a fake made with AI. The problem was the timing. Slovak election law bans election-related reporting during the 48 hours before voting. It is the so-called election silence period. This law, a relic of the old media era, worked in reverse in the social media era. Even if established media conducted fact-checks, they could not officially report them, and the fake spread freely on Facebook.

There is an even more uncomfortable part. On the same day, Russia's Foreign Intelligence Service (SVR) issued a statement claiming that "the United States is interfering in Slovakia's election and backing Progressive Slovakia." Russian state broadcaster Rossija 1 gave major coverage to this story in its election-eve news. The background and foreground moved in the same direction.

Šimečka, who had been ahead in the polls, lost. The pro-Russian Robert Fico won, and immediately after being elected, he pledged to stop military aid to Ukraine and oppose sanctions on Russia. Harvard's HKS Misinformation Review defined this election as the 'Slovak case' that gave rise to the interpretation that it was "the first election whose result was changed by a deepfake."

Another point to remember: Meta's manipulated-content rules at the time covered only video deepfakes. Audio deepfakes were outside the rules. Denník N asked for removal, but Meta did not even respond. It was clear proof that the rules had failed to keep up with the technology.

India 2024: Sixty Million AI-Cloned Voice Calls

By scale, India comes first. India's general election, held from April to June 2024, was the world's largest election, with 968 million voters participating. AI came to the front in this election. According to the Harvard Political Review, in the two months before the election, more than 50 million calls using AI-cloned voices were made. Industry estimates say a new market worth $60 million opened.

The uses split in two directions: positive use and abuse. Start with the positive side. A representative case was Prime Minister Modi using AI to translate a Hindi speech into Tamil, Bengali, and Telugu in real time and deliver it to southern voters. Personalized video messages beginning with "Hello, Hong Gil-dong" were sent in bulk to residents of remote villages who could not attend rallies. There were also cases where deceased senior politicians were brought back as holograms to make statements of support.

The abuse side was serious. A video swept social media showing Rahul Gandhi of the opposition Indian National Congress declaring, "I can no longer pretend to be a Hindu for elections," and leaving the party. The video was a synthetic work that overlaid an AI-cloned voice onto real footage of a lawmaker's oath. Videos of Bollywood stars Aamir Khan and Ranveer Singh criticizing Prime Minister Modi and appealing for support for the opposition were also viewed more than 500,000 times. They were all fakes made by AI.

India's Election Commission sent warning letters to political parties, but there were few regulatory provisions. Divyendra Singh Jadoun, who calls himself the "Indian Deepfaker," testified that "this was the first time AI was sprayed across an election at this scale." He disclosed cases where he refused requests to make fake videos attacking opposing candidates. There was money in it, but he did not take it. Many others did.

New Hampshire 2024: Biden's Voice Stopped Voting

The United States could not avoid it either. In January 2024, on the weekend before the New Hampshire Democratic primary, phones rang in thousands of households across the state. The caller ID showed the name of a local voter, and from the receiver came President Joe Biden's voice. The message was short: "Do not vote now. Save it for the November general election."

Biden had never said that. It was an AI-cloned voice. It was recorded as the first nationwide use of a deepfake in U.S. political history. The investigation moved quickly. New Hampshire prosecutors identified Democratic political consultant Steve Kramer. He had been working for the campaign of Representative Dean Phillips, a candidate in the same party's primary. The Phillips campaign immediately stated that it had nothing to do with them.

The more shocking part was how it was made. Kramer paid a magician in New Orleans $150 to create the voice. The magician told NBC that production took a little over 20 minutes and that the direct cost of making the fake voice was $1. A world where one dollar can create the president's voice and $150 can shake the voting intentions of thousands had actually opened.

The punishment was heavy. The U.S. Federal Communications Commission (FCC) imposed a $6 million fine on Kramer. Lingo Telecom, the telecom company that transmitted the calls, was fined $1 million. New Hampshire charged Kramer with 26 criminal offenses in total, including 13 counts of voter suppression and 13 counts of impersonating a candidate. Soon afterward, the FCC created a new rule that broadly bans AI voice robocalls without consent.

Pushpaganda: Fake News Planted in Google Feeds

In April 2026, a new method appeared. It was a campaign named "Pushpaganda" by the security outlet The Hacker News. The stage was the Google Discover feed that appears on the home screens of Android phones and Chrome browser users. In the place that automatically recommends articles a user might like, AI-made fear-inducing fake articles were inserted.

The method had two steps. Articles were made in a structure search engines like, exposed near the top, and then mass-produced by AI. The goal was to siphon advertising revenue, but the harm comes to users. Fake information appears on an ordinary smartphone screen like a legitimate article. If the methods in Slovakia, India, and New Hampshire were ways of "directly sending fake audio or video," Pushpaganda is one level higher. It uses the platform's personalized recommendation algorithm itself as the channel.

Slovakia, India, New Hampshire, the USC experiment, and Pushpaganda. These five scenes belong to the same single-line narrative: the side creating fakes needs only $1 and 20 minutes, while the side catching them must rebuild law, rules, and technology. It is an asymmetry.

Think about a presidential election. Suppose that two days before voting, an audio file of a major candidate spreads through Telegram. Its content is "a scene discussing money-for-nominations for district candidacies." Fact-checks will of course follow. But what Slovakia showed was that fakes spread much faster than fact-checks can attach themselves. What India showed was that this method can be sprayed to tens of millions of people one phone call at a time. What New Hampshire showed was that the production cost of $150 is enough.

Korea's current Public Official Election Act bans deepfake election campaigning from 90 days before election day (Article 82-8). This provision was newly created in January 2024. But it does not fully cover voice cloning, AI automated calls, or infiltration of recommendation algorithms. The point indicated by the USC study is more painful. When AI agents coordinate with one another to create public opinion, there is no human manipulator, so the existing law's provisions on responsible actors themselves do not work. The law was made to punish people, but what should be done when the actor that manipulated is not a person? This is the homework of 2026.

AI in the Security Field: Anthropic Mythos Preview

Around the same time, AI is active on another battlefield as well. On April 7, 2026, Anthropic released a model called Claude Mythos Preview. Over the past few weeks, this model has found thousands of security holes, or zero-days, on its own across Windows, Mac, Linux, Chrome, and Safari. In OpenBSD, it found a 27-year-old bug, and in FFmpeg, a 16-year-old bug. More than 99% of the discovered vulnerabilities remain unpatched.

Anthropic decided not to release this model to the public. Instead, under the name Project Glasswing, it formed a defense-only consortium with 12 companies including Amazon, Apple, Google, Microsoft, and NVIDIA. $100 million in model-use credits will be committed. OpenAI soon followed by releasing a security-specialized model called GPT-5.4-Cyber only to verified defenders. A system in which high-performance defensive AI is not released to the public but goes only to verified companies is becoming the new standard of spring 2026.

China's Response: Alibaba Token Hub and the Zhenwu Chip

China moved too. On March 16, Alibaba merged five AI-related organizations into Alibaba Token Hub(ATH)under that name. CEO Eddie Wu will lead it personally. Over three years, 380 billion yuan, about KRW 73 trillion, will be invested. On April 8, it opened a data center in southern China running only on 10,000 self-designed Zhenwu (真武) AI chips. It is a declaration that it will operate AI data centers without NVIDIA GPUs.

In the AI model marketplace OpenRouter, the usage share of Chinese models surpassed that of U.S. models in spring 2026. Moonshot, the developer of Kimi, and MiniMax are pursuing listings on the Hong Kong Stock Exchange. Moonshot's valuation is around $18 billion. When Korean companies actually spend money on AI services, whether to include Chinese models is no longer an ideology issue, but a matter of unit cost and performance.

If Threaded into One Line

In the end, all of today's scenes are threaded onto the same line. AI creates public opinion, finds security holes, and rewrites price tags. Offensive capability and defensive capability grow from the same technology, and the time gap between them is getting shorter. In Slovakia it was two days, in New Hampshire one weekend day, and in the USC experiment, "no person was present."

The tasks can be arranged into three: tightly revising election law, codifying the anomaly-detection responsibilities of platforms such as Meta, X, and Kakao, and building national-level verification infrastructure that lets citizens judge within 30 seconds that "this video is highly likely to have been made by AI". The last task is the heaviest. Law follows after the fact, but citizens' eyes must work in real time.