Protected: News

In January 2024, the phone rang in homes all around New Hampshire. On the other end was Joe Biden’s voice, urging Democrats to “save your vote” by skipping the primary. It sounded authentic, but it wasn’t. The call was a fake, generated by artificial intelligence.

Today, the technology behind that hoax looks quaint. Tools like OpenAI’s Sora now make it possible to create convincing synthetic videos with astonishing ease. AI can be used to fabricate messages from politicians and celebrities—even entire news clips—in minutes. The fear that elections could be overwhelmed by realistic fake media has gone mainstream—and for good reason.

But that’s only half the story. The deeper threat isn’t that AI can just imitate people—it’s that it can actively persuade people. And new research published this week shows just how powerful that persuasion can be. In two large peer-reviewed studies, AI chatbots shifted voters’ views by a substantial margin, far more than traditional political advertising tends to do.

In the coming years, we will see the rise of AI that can personalize arguments, test what works, and quietly reshape political views at scale. That shift—from imitation to active persuasion—should worry us deeply.  

The challenge is that modern AI doesn’t just copy voices or faces; it holds conversations, reads emotions, and tailors its tone to persuade. And it can now command other AIs—directing image, video, and voice models to generate the most convincing content for each target. Putting these pieces together, it’s not hard to imagine how one could build a coordinated persuasion machine. One AI might write the message, another could create the visuals, another could distribute it across platforms and watch what works. No humans required.

A decade ago, mounting an effective online influence campaign typically meant deploying armies of people running fake accounts and meme farms. Now that kind of work can be automated—cheaply and invisibly. The same technology that powers customer service bots and tutoring apps can be repurposed to nudge political opinions or amplify a government’s preferred narrative. And the persuasion doesn’t have to be confined to ads or robocalls. It can be woven into the tools people already use every day—social media feeds, language learning apps, dating platforms, or even voice assistants built and sold by parties trying to influence the American public. That kind of influence could come from malicious actors using the APIs of popular AI tools people already rely on, or from entirely new apps built with the persuasion baked in from the start.

And it’s affordable. For less than a million dollars, anyone can generate personalized, conversational messages for every registered voter in America. The math isn’t complicated. Assume 10 brief exchanges per person—around 2,700 tokens of text—and price them at current rates for ChatGPT’s API. Even with a population of 174 million registered voters, the total still comes in under $1 million. The 80,000 swing voters who decided the 2016 election could be targeted for less than $3,000. 

Although this is a challenge in elections across the world, the stakes for the United States are especially high, given the scale of its elections and the attention they attract from foreign actors. If the US doesn’t move fast, the next presidential election in 2028, or even the midterms in 2026, could be won by whoever automates persuasion first. 

The 2028 threat 

While there have been indications that the threat AI poses to elections is overblown, a growing body of research suggests the situation could be changing. Recent studies have shown that GPT-4 can exceed the persuasive capabilities of communications experts when generating statements on polarizing US political topics, and it is more persuasive than non-expert humans two-thirds of the time when debating real voters. 

Two major studies published yesterday extend those findings to real election contexts in the United States, Canada, Poland, and the United Kingdom, showing that brief chatbot conversations can move voters’ attitudes by up to 10 percentage points, with US participant opinions shifting nearly four times more than it did in response to tested 2016 and 2020 political ads. And when models were explicitly optimized for persuasion, the shift soared to 25 percentage points—an almost unfathomable difference.

While previously confined to well-resourced companies, modern large language models are becoming increasingly easy to use. Major AI providers like OpenAI, Anthropic, and Google wrap their frontier models in usage policies, automated safety filters, and account-level monitoring, and they do sometimes suspend users who violate those rules. But those restrictions apply only to traffic that goes through their platforms; they don’t extend to the rapidly growing ecosystem of open-source and open-weight models, which  can be downloaded by anyone with an internet connection. Though they’re usually smaller and less capable than their commercial counterparts, research has shown with careful prompting and fine-tuning, these models can now match the performance of leading commercial systems. 

All this means that actors, whether well-resourced organizations or grassroots collectives, have a clear path to deploying politically persuasive AI at scale. Early demonstrations have already occurred elsewhere in the world. In India’s 2024 general election, tens of millions of dollars were reportedly spent on AI to segment voters, identify swing voters, deliver personalized messaging through robocalls and chatbots, and more. In Taiwan, officials and researchers have documented China-linked operations using generative AI to produce more subtle disinformation, ranging from deepfakes to language model outputs that are biased toward messaging approved by the Chinese Communist Party.

It’s only a matter of time before this technology comes to US elections—if it hasn’t already. Foreign adversaries are well positioned to move first. China, Russia, Iran, and others already maintain networks of troll farms, bot accounts, and covert influence operators. Paired with open-source language models that generate fluent and localized political content, those operations can be supercharged. In fact, there is no longer a need for human operators who understand the language or the context. With light tuning, a model can impersonate a neighborhood organizer, a union rep, or a disaffected parent without a person ever setting foot in the country. Political campaigns themselves will likely be close behind. Every major operation already segments voters, tests messages, and optimizes delivery. AI lowers the cost of doing all that. Instead of poll-testing a slogan, a campaign can generate hundreds of arguments, deliver them one on one, and watch in real time which ones shift opinions.

The underlying fact is simple: Persuasion has become effective and cheap. Campaigns, PACs, foreign actors, advocacy groups, and opportunists are all playing on the same field—and there are very few rules.

The policy vacuum

Most policymakers have not caught up. Over the past several years, legislators in the US have focused on deepfakes but have ignored the wider persuasive threat.

Foreign governments have begun to take the problem more seriously. The European Union’s 2024 AI Act classifies election-related persuasion as a “high-risk” use case. Any system designed to influence voting behavior is now subject to strict requirements. Administrative tools, like AI systems used to plan campaign events or optimize logistics, are exempt. However, tools that aim to shape political beliefs or voting decisions are not.

By contrast, the United States has so far refused to draw any meaningful lines. There are no binding rules about what constitutes a political influence operation, no external standards to guide enforcement, and no shared infrastructure for tracking AI-generated persuasion across platforms. The federal and state governments have gestured toward regulation—the Federal Election Commission is applying old fraud provisions, the Federal Communications Commission has proposed narrow disclosure rules for broadcast ads, and a handful of states have passed deepfake laws—but these efforts are piecemeal and leave most digital campaigning untouched. 

In practice, the responsibility for detecting and dismantling covert campaigns has been left almost entirely to private companies, each with its own rules, incentives, and blind spots. Google and Meta have adopted policies requiring disclosure when political ads are generated using AI. X has remained largely silent on this, while TikTok bans all paid political advertising. However, these rules, modest as they are, cover only the sliver of content that is bought and publicly displayed. They say almost nothing about the unpaid, private persuasion campaigns that may matter most.

To their credit, some firms have begun publishing periodic threat reports identifying covert influence campaigns. Anthropic, OpenAI, Meta, and Google have all disclosed takedowns of inauthentic accounts. However, these efforts are voluntary and not subject to independent auditing. Most important, none of this prevents determined actors from bypassing platform restrictions altogether with open-source models and off-platform infrastructure.

What a real strategy would look like

The United States does not need to ban AI from political life. Some applications may even strengthen democracy. A well-designed candidate chatbot could help voters understand where the candidate stands on key issues, answer questions directly, or translate complex policy into plain language. Research has even shown that AI can reduce belief in conspiracy theories. 

Still, there are a few things the United States should do to protect against the threat of AI persuasion. First, it must guard against foreign-made political technology with built-in persuasion capabilities. Adversarial political technology could take the form of a foreign-produced video game where in-game characters echo political talking points, a social media platform whose recommendation algorithm tilts toward certain narratives, or a language learning app that slips subtle messages into daily lessons. Evaluations, such as the Center for AI Standards and Innovation’s recent analysis of DeepSeek, should focus on identifying and assessing AI products—particularly from countries like China, Russia, or Iran—before they are widely deployed. This effort would require coordination among intelligence agencies, regulators, and platforms to spot and address risks.

Second, the United States should lead in shaping the rules around AI-driven persuasion. That includes tightening access to computing power for large-scale foreign persuasion efforts, since many actors will either rent existing models or lease the GPU capacity to train their own. It also means establishing clear technical standards—through governments, standards bodies, and voluntary industry commitments—for how AI systems capable of generating political content should operate, especially during sensitive election periods. And domestically, the United States needs to determine what kinds of disclosures should apply to AI-generated political messaging while navigating First Amendment concerns.

Finally, foreign adversaries will try to evade these safeguards—using offshore servers, open-source models, or intermediaries in third countries. That is why the United States also needs a foreign policy response. Multilateral election integrity agreements should codify a basic norm: States that deploy AI systems to manipulate another country’s electorate risk coordinated sanctions and public exposure. 

Doing so will likely involve building shared monitoring infrastructure, aligning disclosure and provenance standards, and being prepared to conduct coordinated takedowns of cross-border persuasion campaigns—because many of these operations are already moving into opaque spaces where our current detection tools are weak. The US should also push to make election manipulation part of the broader agenda at forums like the G7 and OECD, ensuring that threats related to AI persuasion are treated not as isolated tech problems but as collective security challenges.

Indeed, the task of securing elections cannot fall to the United States alone. A functioning radar system for AI persuasion will require partnerships with our partners and allies. Influence campaigns are rarely confined by borders, and open-source models and offshore servers will always exist. The goal is not to eliminate them but to raise the cost of misuse and shrink the window in which they can operate undetected across jurisdictions.

The era of AI persuasion is just around the corner, and America’s adversaries are prepared. In the US, on the other hand, the laws are out of date, the guardrails too narrow, and the oversight largely voluntary. If the last decade was shaped by viral lies and doctored videos, the next will be shaped by a subtler force: messages that sound reasonable, familiar, and just persuasive enough to change hearts and minds.

For China, Russia, Iran, and others, exploiting America’s open information ecosystem is a strategic opportunity. We need a strategy that treats AI persuasion not as a distant threat but as a present fact. That means soberly assessing the risks to democratic discourse, putting real standards in place, and building a technical and legal infrastructure around them. Because if we wait until we can see it happening, it will already be too late.

Tal Feldman is a JD candidate at Yale Law School who focuses on technology and national security. Before law school, he built AI models across the federal government and was a Schwarzman and Truman scholar. Aneesh Pappu is a PhD student and Knight-Hennessy scholar at Stanford University who focuses on agentic AI and technology policy. Before Stanford, he was a privacy and security researcher at Google DeepMind and a Marshall scholar. 

In 2024, a Democratic congressional candidate in Pennsylvania, Shamaine Daniels, used an AI chatbot named Ashley to call voters and carry on conversations with them. “Hello. My name is Ashley, and I’m an artificial intelligence volunteer for Shamaine Daniels’s run for Congress,” the calls began. Daniels didn’t ultimately win. But maybe those calls helped her cause: New research reveals that AI chatbots can shift voters’ opinions in a single conversation—and they’re surprisingly good at it. 

A multi-university team of researchers has found that chatting with a politically biased AI model was more effective than political advertisements at nudging both Democrats and Republicans to support presidential candidates of the opposing party. The chatbots swayed opinions by citing facts and evidence, but they were not always accurate—in fact, the researchers found, the most persuasive models said the most untrue things. 

The findings, detailed in a pair of studies published in the journals Nature and Science, are the latest in an emerging body of research demonstrating the persuasive power of LLMs. They raise profound questions about how generative AI could reshape elections. 

“One conversation with an LLM has a pretty meaningful effect on salient election choices,” says Gordon Pennycook, a psychologist at Cornell University who worked on the Nature study. LLMs can persuade people more effectively than political advertisements because they generate much more information in real time and strategically deploy it in conversations, he says. 

For the Nature paper, the researchers recruited more than 2,300 participants to engage in a conversation with a chatbot two months before the 2024 US presidential election. The chatbot, which was trained to advocate for either one of the top two candidates, was surprisingly persuasive, especially when discussing candidates’ policy platforms on issues such as the economy and health care. Donald Trump supporters who chatted with an AI model favoring Kamala Harris became slightly more inclined to support Harris, moving 3.9 points toward her on a 100-point scale. That was roughly four times the measured effect of political advertisements during the 2016 and 2020 elections. The AI model favoring Trump moved Harris supporters 2.3 points toward Trump. 

In similar experiments conducted during the lead-ups to the 2025 Canadian federal election and the 2025 Polish presidential election, the team found an even larger effect. The chatbots shifted opposition voters’ attitudes by about 10 points.

Long-standing theories of politically motivated reasoning hold that partisan voters are impervious to facts and evidence that contradict their beliefs. But the researchers found that the chatbots, which used a range of models including variants of GPT and DeepSeek, were more persuasive when they were instructed to use facts and evidence than when they were told not to do so. “People are updating on the basis of the facts and information that the model is providing to them,” says Thomas Costello, a psychologist at American University, who worked on the project. 

The catch is, some of the “evidence” and “facts” the chatbots presented were untrue. Across all three countries, chatbots advocating for right-leaning candidates made a larger number of inaccurate claims than those advocating for left-leaning candidates. The underlying models are trained on vast amounts of human-written text, which means they reproduce real-world phenomena—including “political communication that comes from the right, which tends to be less accurate,” according to studies of partisan social media posts, says Costello.

In the other study published this week, in Science, an overlapping team of researchers investigated what makes these chatbots so persuasive. They deployed 19 LLMs to interact with nearly 77,000 participants from the UK on more than 700 political issues while varying factors like computational power, training techniques, and rhetorical strategies. 

The most effective way to make the models persuasive was to instruct them to pack their arguments with facts and evidence and then give them additional training by feeding them examples of persuasive conversations. In fact, the most persuasive model shifted participants who initially disagreed with a political statement 26.1 points toward agreeing. “These are really large treatment effects,” says Kobi Hackenburg, a research scientist at the UK AI Security Institute, who worked on the project. 

But optimizing persuasiveness came at the cost of truthfulness. When the models became more persuasive, they increasingly provided misleading or false information—and no one is sure why. “It could be that as the models learn to deploy more and more facts, they essentially reach to the bottom of the barrel of stuff they know, so the facts get worse-quality,” says Hackenburg.

The chatbots’ persuasive power could have profound consequences for the future of democracy, the authors note. Political campaigns that use AI chatbots could shape public opinion in ways that compromise voters’ ability to make independent political judgments.

Still, the exact contours of the impact remain to be seen. “We’re not sure what future campaigns might look like and how they might incorporate these kinds of technologies,” says Andy Guess, a political scientist at Princeton University. Competing for voters’ attention is expensive and difficult, and getting them to engage in long political conversations with chatbots might be challenging. “Is this going to be the way that people inform themselves about politics, or is this going to be more of a niche activity?” he asks.

Even if chatbots do become a bigger part of elections, it’s not clear whether they’ll do more to  amplify truth or fiction. Usually, misinformation has an informational advantage in a campaign, so the emergence of electioneering AIs “might mean we’re headed for a disaster,” says Alex Coppock, a political scientist at Northwestern University. “But it’s also possible that means that now, correct information will also be scalable.”

And then the question is who will have the upper hand. “If everybody has their chatbots running around in the wild, does that mean that we’ll just persuade ourselves to a draw?” Coppock asks. But there are reasons to doubt a stalemate. Politicians’ access to the most persuasive models may not be evenly distributed. And voters across the political spectrum may have different levels of engagement with chatbots. “If supporters of one candidate or party are more tech savvy than the other,” the persuasive impacts might not balance out, says Guess.

As people turn to AI to help them navigate their lives, they may also start asking chatbots for voting advice whether campaigns prompt the interaction or not. That may be a troubling world for democracy, unless there are strong guardrails to keep the systems in check. Auditing and documenting the accuracy of LLM outputs in conversations about politics may be a first step.