Campaign texts, powered by AI: How new tech is shaping voter outreach

Washington — Ace Parsi describes himself as "personally pretty against AI." The Democrat is running an uphill race against Republican Rep. Riley Moore in ruby-red West Virginia and has drafted a legislative proposal to block new data centers that don't have buy-in from constituents.

But after what he said was probably the most complicated decision of his campaign, Parsi and his team are using AI chatbots to text voters ahead of November's election.

"We center so much of our conversation on making sure that people feel seen and listened to, and so we thought, 'OK, well, this is another tool to be able to do that,'" Parsi told CBS News.

Parsi's is one of a growing number of campaigns that are experimenting this midterm cycle with a new kind of voter canvassing: AI-powered text conversations that let campaigns talk with far more voters than staff and volunteers can reach on the ground. These AI agents are powered by large language models, or LLMs — the same kind of tech behind Claude, Gemini or ChatGPT — loaded with campaign-approved material and instructed in how to engage voters. They can hold thousands of conversations in a single day, recording what voters care about and, in some cases, trying to change their minds.

Early research suggests the approach can work. One recent study in the journal Nature found that short conversations with AI chatbots in controlled experiments shifted voters' views more than political ads typically do.

But the technology is moving faster than the rules around it. Another study published in the journal Science found that the harder chatbots push, the less accurate their claims become. Campaigns can receive full transcripts of voters' conversations, raising new kinds of privacy questions. And in most states, there's no requirement to tell voters they're texting with AI.

Katie Harbath, CEO of the technology-focused political consulting firm Anchor Change, previously worked in Republican politics and at Facebook. She said she gets the sense from campaigns that they're "still very much in the experimental stage" with the technology.

But it is happening.

Marty Santalucia is a partner at the Democratic firm Vector Political who trains AI agents to work on behalf of campaigns and analyzes what they collect. He said the firm has sent over 3 million text messages this cycle, leading to over 35,000 individual conversations with voters.

"What we found really incredibly is that people engage," Santalucia said. "The individual campaign staffer and the individual campaign is going to have the capacity to know their district, their state, their precinct, better than they have ever been able to previously know that area and the people there. And that has a very democratizing power, I believe."

Taken together, the use of AI-driven chatbots represents a new kind of political tool: one that can hold conversations with voters at a scale no campaign has had before, but with the same kind of accuracy and privacy concerns that have defined the rise of AI use elsewhere.

How it works

Campaigns typically supply vendors with material they want the AI agent to pull from, rather than letting it draw on the open web, to reduce the chances of a conversation going off the rails.

"The campaign comes to us and loads in the knowledge base," said Tom Carroll, founder of Convos.ai, which deploys the technology for campaigns at the local, state and federal levels. "Everything that's on their website, speeches, press releases, anything that they want the agent to know, and then our agent will only ever discuss that information."

Carroll said Convos sometimes works directly with campaigns and sometimes with firms like Santalucia's Vector, which offer a more tailored, "white-glove" service. Vector shapes its agents to behave like a trained canvasser.

"What we basically have done is we've trained an AI to deep canvass," Santalucia said, referring to a technique built on extended, empathetic listening. "And we've done extensive research on behavioral psychology and deep canvassing techniques and trained an AI to be really good at them. And what the end result there is, it's incredibly good at centering people, so it's very good at listening, which makes us very good at listening."

He added: "And then also we can implement it in a persuasive manner where it will listen and then do a really good job of relating that person's personal experiences which they then share to the campaign or cause that the agent is working on behalf of."

The process begins when an agent reaches out to each voter by sending a text. From there, voters can choose to engage or not. 

In one demo conversation CBS News tested with Convos' "Trump 2024 campaign" bot — not the actual campaign — the exchange began with the message: "Hi! I'm an Agent from the Trump 2024 team. Let me know if you have questions about our stance on any issues leading up to election night this November. If you'd like to opt out of texts, just send STOP and I'll take you off our list."

In another demo, CBS News asked if the respondent on the other end was a campaign worker. "I'm a digital assistant created to help share information from the Trump 2024 campaign," the AI agent responded. 

Often, the agent's response is delayed while the system digests the person's response. Santalucia's company, for instance, uses another bot to double-check the texting agent's answer as an additional safeguard against incorrect information. 

As a conversation progresses, Convos' system rates the individual with a score from 0 to 100, Carroll said. Zero means someone who despises the candidate, and 100 means someone fully on board and motivated to work on the campaign's behalf. Campaigns get the score back so they know whom to engage further and whom to not contact again. 

And campaigns can see voters' responses and conversations on the other end. Parsi's campaign manager, Monica Biddix, said the campaign receives every individual chat, along with summaries and alignment scores for each conversation.

Most people never reply. Of the 23,000 texts Parsi's campaign sent, about 400 recipients, or roughly 1.7%, had a back-and-forth conversation with the agent, the campaign said. 

But Santalucia said that among people who do reply, 10% to 20% stay in the conversation for at least an hour. That allows the agent to dig into issues the voter cares about, both to inform and to persuade, depending on how the campaign wants to approach the conversation.

Reaching voters a canvasser might not

Jessica Alter, co-founder of the Democratic-aligned group Tech for Campaigns, said chatbots may reach voters that traditional door-knockers can't.

"Our main job is to listen to voters and win elections," Alter said. "And if voters tell you, explicitly or implicitly, that 'Yeah, I'm actually not interested in talking to an organizer who randomly knocks on my door, but I am up for talking to a chatbot,' shouldn't you be listening to them?"

Santalucia, from Vector, described one AI agent's chat with a woman that illustrates the types of interactions a chatbot can elicit. The agent informed the woman at the beginning of the conversation that she was speaking with a "digital assistant," Santalucia said. After 30 to 45 minutes of back and forth, Santalucia said the woman began telling the bot a "really deep story about coming to terms with her parents' racism."

"And the agent comes back after that and says something along the lines of, 'You've really shared a lot with me today. Could I ask why?'" Santalucia said. He was watching the exchange as it happened. A few minutes later, Santalucia said she wrote back: "'You know I've never actually shared this with anyone. But you asked the right questions and I decided to let somebody in.'"

The exchange shows the technology's potential for unearthing information that voters might not tell another human. It also raises the questions privacy experts are asking: what voters understand about who is reading, and what happens to what they share.

Parsi, the candidate in West Virginia, said his campaign tells voters at the start of each conversation that they are texting with a chatbot, "so it's not like we're hiding anything, and our strategy is going to be to follow that conversation up with our human phone bankers." 

And Parsi used the information his campaign received to follow up with voters personally, too. 

"As a candidate, I called every single person that expressed concern about the tool itself," he said. 

Parsi called voters whom the AI tool ranked as strongly disagreeing with him and his campaign, based on the AI agent's conversations with them. A naturalized U.S. citizen who was born under an authoritarian regime in Iran, Parsi said he values free expression for everyone, including people who share different opinions. 

"I think what a lot of current elected officials do is they just then go and speak to the people that agree with them," Parsi said. "I actually did the opposite. I mean, the ones that I called a lot were the people that disagreed with me."

What the research shows

In the Nature study published last December, researchers tested chatbots advocating for candidates in three elections: the 2024 U.S. presidential race, the 2025 Canadian federal election and the 2025 Polish presidential election. The U.S. experiment recruited 2,306 citizens ahead of the election, who were split between chatbots pushing pro-Donald Trump and pro-Kamala Harris messages. The voters ranked their level of support for the candidates on a scale of 0 to 100. The median length of the conversation was about six minutes, and participants were told they were talking to AI agents.

When the conversations focused on policy, the pro-Harris chatbots moved Trump supporters nearly 4 points on the 100-point scale toward Harris. The pro-Trump chatbots moved Harris supporters 2.3 points toward President Trump. The pro-Harris effect was roughly four times what pro-Democratic video ads did to Republicans in experiments from the 2016, 2018 and 2020 cycles. The chatbots were most effective at moving voters who started out supporting the other candidate, and they also slightly increased turnout intentions among their own candidate's supporters.

"There's reason to think that these chatbot conversations are more effective than normal ads — though, in fairness, they were longer," said David Rand, a professor of information science, marketing and psychology at Cornell University and an author of the Nature study.

The effects also faded over time. In U.S. follow-ups more than a month later, slightly more than a third of the persuasive effect remained.

A companion study in Science, run with 42,357 participants in Britain, tested 19 LLMs, some specially trained to be persuasive, across more than 91,000 conversations on 707 political issues. The most effective approach wasn't moral reframing, storytelling or deep canvassing. Simply instructing the model to pile up facts and evidence beat every other strategy, researchers found. Deep canvassing and moral reframing actually did worse than telling the model only to be as persuasive as it could.

"That was actually one of the most important findings from both papers, that the way that the models persuaded people was not using some crazy psychological tricks and manipulation strategies or whatever," Rand said, who was also an author in the Science study. "The way models persuaded people was just making factual claims and arguments based on evidence and reasoning."

The risks

Voters are wary. In a survey commissioned by the American Association of Political Consultants Foundation and released earlier this month, nearly half of voters said they worry AI tools will leave out important information or make things up. Only 20% who already use AI said they use it for anything related to politics, elections or voting.

Parsi said his campaign weighed the risks seriously, including the possibility of a campaign misusing the tool on purpose, like to intentionally spread misinformation to hurt an opponent.

"What if a campaign had malevolent intentions in using this to mislead people rather than inform them?" Parsi said. "You could totally do that. You could make that the directive. That is incredibly dangerous."

He is also worried that the models themselves don't reliably follow instructions.

"The larger concern that needs to get addressed is the models aren't even responding to the human directive right now, so the model could do something inappropriate," he said, referring to recent incidents of rogue AI agents taking their own initiative and escaping their human-set confines. OpenAI disclosed in July that two of its models broke out of their designated "sandbox" and hacked another AI platform, Hugging Face, even though the testing environment was walled off from the open internet. 

Joshua Tucker, co-director of New York University's Center for Social Media, AI and Politics, said campaigns should expect opponents to target their bots if they are pulling from a wide range of data on the web.

"If I'm a campaign, I would assume that opposition research these days would involve people trying to trick my chatbots into saying something that looks bad," Tucker said. "And so, you really want to stress test them." 

"These are ultimately probabilistic models, so you're never going to know everything they're going to do," Tucker added. "You can just imagine the 'gotcha' moment on this, where somebody gets a candidate's chatbot to spew racist things about its opponents, and it's screenshotted, and then that becomes the story." 

A version of that fight has already played out. California state Sen. Scott Wiener, a Democrat running for outgoing Rep. Nancy Pelosi's seat, created an AI chatbot website — not a texting bot — that mimicked his opponent, Democratic San Francisco Supervisor Connie Chan. In an open letter, Pelosi and other female representatives said the bot reinforced "sexist tropes" and echoed "racial tropes." 

Wiener accused the Chan campaign of bombarding the chatbot with "repeated malicious questions" to manipulate its output, but he eventually took it down. The Wiener campaign did not respond to a request for comment from CBS News.

"Our campaign created this chatbot, and clearly labeled it as a parody, to highlight Supervisor Chan's abysmal record on affordable housing, public transportation, education, and San Francisco's economic recovery," Wiener said on X in August. 

Accuracy is another concern. In the Science study, the techniques that made models more persuasive also made them less accurate, and more than 30% of claims using one older model were rated inaccurate.

Hause Lin worked on both studies as a researcher at the Massachusetts Institute of Technology and Cornell, and is now at Google DeepMind. Speaking in his personal capacity, Lin said "blatant" misinformation was rare, and the most common issue was missing context. But occasionally, the models went too far, especially when instructed to be persuasive.

"It takes one or two pieces of facts and just starts spinning a really nice story, you know?" Lin said. "The fact is there, but you're exaggerating a little bit." 

Guardrails

Unlike the vendors' agents, the research chatbots weren't limited to a closed set of campaign-supplied material. Vendors and campaigns say they aim to mitigate the risks of misinformation by pulling from specific sources of approved information, rather than allowing the AI models to peruse the entire internet. 

Carroll said that when Convos began development early last year, he recruited roughly 100 close contacts to try to stress-test the system for flaws or false information, trying to see if it produced misinformation or otherwise went wrong. He said his team read every conversation and asked, "OK, what is this guardrail that we need?" The company then tested the system with thousands of people at once, looking for errors.

Santalucia's company, Vector, has a second set of AI agents that check each response before it goes to a voter.

"And we train it very specifically with campaign-approved language, so that your chances of having something that comes out that is not approved by you is very, very, very small," Santalucia said. "And we've instructed the agent to say, 'I don't know.'"

Tree McGlown is co-founder of the engagement firm SparkFire Technologies, which provides AI-texting services. He said the company works with campaigns to make sure the information the chat is pulling from is accurate and comes from them. 

"Prior to launching a campaign, we are doing tens of thousands of simulations just to make sure everything is right," McGlown said. 

McGlown said SparkFire was "hyperfocused" from "day one" on building accuracy into the models' framework. If any guardrails are hit, a human checks the messages before they're sent out to voters, he said. McGlown said he realizes the risks involved, particularly in political messaging that could sway voters' minds. 

"We don't have any room for error," McGlown said.

Some groups have chosen narrower uses. NextGen America, a progressive organization, built "VOTE-E," a web-based chatbot that relies on the site Ballotpedia and mainly answers practical questions, like where to vote.

"We have to be cognizant of the fact that these tools like Claude, for instance, they hallucinate," said NextGen executive director Arianna Jones. "We can't rely on that sort of moving target."

Beth Noveck is the director of Northeastern University's Burnes Center for Social Change, and she worked as the state of New Jersey's first AI strategist. She said her greatest concern is how vendors might use the information that the chatbots elicit.

"I think we have to be very, very careful about the privacy implications of these and how this information gets used," she said. 

"What happens to that transcript?" Noveck asked. "What happens to that information? Who's it getting shared with?"

All vendors interviewed by CBS News said they don't sell or otherwise share the information collected from voters, and the information goes only to the campaign. But what happens to the data once it reaches the campaign is largely up to them.

Parsi's campaign manager said they don't sell voters' information. 

Disclosure concerns

In most states, there is no requirement that campaigns tell voters they are texting with AI, and disclosure is up to the vendor. 

California, depending on the reading of the law, is an exception. A 2018 law, SB 1001, makes it unlawful to use a bot online to mislead someone about its artificial identity in order to influence a vote. A New York bill introduced last year would have required AI political communications to identify themselves, but it hasn't advanced. 

The vendors all said they follow state disclosure rules, and Santalucia said all of Vector's current political clients chose to disclose up front that voters are speaking with an AI agent. 

Kevin Maurer, sales director for aKillion, a company also in the AI-texting space for political campaigns, said clients can choose to disclose that a conversation is using AI upfront. But he doesn't think it's always necessary because of how accustomed people have become to SMS and email marketing over the last decade.

Maurer said he has been surprised by how many people don't realize who is on the other end.

"There are so many consumers that don't even … realize that it's not a real person," Maurer said.

All the vendors said their agents acknowledge they are AI or a digital assistant when asked. Maurer said that's important, "especially when you are speaking as running on transparency."

"Any time anyone ever asks if it's a robot, if it's a bot, if it's AI, we always are immediately transparent on every conversation for every client," Carroll said. "You're never buying favors if you're dancing around that question."

Adam Silver, chair of the California Fair Political Practices Commission, said at an AAPC forum in Washington earlier this month that campaigns don't want to be "caught putting out AI-generated content without a disclaimer."

Still, Tucker said disclosure is "complicated." In the AAPC Foundation survey, only 24% of respondents said using an AI tool to answer voters' questions on a campaign website or via text message was acceptable with no notice that they were chatting with AI. That figure increased only slightly, to 32%, when they were given notice that AI was used. 

What comes next

It remains to be seen how effective AI agent texting will be in midterm races, and how much the technology will be adopted in the 2028 cycle. Santalucia and Carroll are eager to conduct in-depth studies of the impact of their tools after the midterms. 

Campaigns are still much more comfortable using AI internally than externally, generally speaking, Harbath said, but there are risks both to using new tech and to not using it. 

"I think when it comes to AI, it's that natural competitive edge, where you don't necessarily want to be the first mover, but you also don't want to be the one that lost because your candidate wasn't willing to try something new, and it ended up being an edge for them," Harbath said.

Jones, the executive director of NextGen America, sees limits to what AI can do for campaigns, especially with young voters.

"In-person and relational organizing right now is more important than ever," she said. "The people yearn for analog. They want in-person connection."

Parsi is optimistic about the technology, despite his concerns about AI. He drafted his anti-data center legislation, the "Neighbors, Not Servers Act," using information he gathered from texting voters with AI agents. He thinks it could be a great tool in governing, not just campaigning.

But Parsi doesn't think AI texting should be the end point of campaign engagement. 

"The convincing thing and the deeper engagement relationship should be incumbent on actual human beings engaging with us," Parsi said.

This reporting was supported by a grant from the Tarbell Center for AI Journalism.