AI is on the verge of changing the practice of campaigning itself

While attention focuses on the rising threat of deep fakes, say hello to AI coaches and AI phone-bots that can replace human volunteers; voters may soon be targeted by hyper-personalized video ads.

Jan 24, 2024

I’m taking a break from the Missile East this week, but I do include some useful readings on the Israel/Palestine quagmire at the bottom of this post. And yes, I haven’t forgotten about the return of the Orange Cheeto, but since everyone else is writing about Iowa and New Hampshire I don’t see a need to pile on. But figuring out how we hold the democratic coalition together as The Most Important Election in Modern American History^TM approaches is still very much front of mind.

Ever since Chat-GPT became the fastest growing new tech platform in Internet history, gaining 100 million users in its first two months, I’ve been expecting artificial intelligence (AI) tools built using large language models (LLMs) to start to hit the political marketplace and transform how campaigns work. Well, that moment is here.

Much of the early attention on AI and campaigns has centered on how synthetic media, also known as “deep fakes,” may increase the supply of disinformation, further weakening democracies and potentially upending elections. And indeed, there have been a number of cases already from around the world where deep fakes have had an impact on particular candidates. Just yesterday, voters in New Hampshire reported receiving a robocall from a voice sounding like Joe Biden, urging people to not vote in the primary. And in New York City, Politico reports that someone used an AI-audio tool to make an ersatz “hot-mic” moment of a Harlem Democratic boss supposedly talking trash about an ailing local Assemblywoman, apparently hoping the leak would create havoc and confusion.

To the degree that legislators and regulators have started to focus on the impact of AI on politics, they’ve responded to these kinds of manifestations by talking about, or in a few states actually enacting, rules that require campaigns to disclose when they use generative AI in political communications, especially in advertising. The “Real Political Advertisements Act” (S. 1596), introduced by Senator Amy Klobuchar, would require that for federal campaigns. California, Michigan, Texas and Minnesota already require such disclosure in the final months or weeks of a campaign, and other states are looking to take similar steps and also give maligned individuals the right to sue. So far, the main response from tech companies is to add digital watermarks to political ads made with AI, to make them easier to identify.

Such moves will only affect how official campaign committees operate; they don’t do anything about this new wave of anonymous dirty tricks. Likewise, OpenAI’s decision to prevent political campaigners from using Chat-GPT (it just blocked supporters of long-shot Democrat Dean Phillips, who had built a Dean.bot chatbot) is no obstacle; with open-source LLMs like Meta’s Llama 2 freely available, a savvy developer has plenty of options.

But while funny/scary videos or visuals with realistic replicas of famous politicians like Presidents Biden, Trump or Obama have circulated widely, showing the public and politicians alike how easy it has suddenly become to put words into their mouths or pose them in compromising positions, this is just the froth at the top of a much bigger wave. AI is on the verge of changing the practice of campaigning itself--how campaigns do everything from strategy to paid advertising to voter outreach. After several years where it seemed that technology’s impact on politics had plateaued, AI-powered tools are about to change the game in some very interesting ways. And like all previous tech waves, this one will not be uniformly good or bad, but nor will its impact be neutral.

Say Hello to Your AI Campaign Assistant

I got a full blast of what’s coming a few days ago, when I got a preview of VotivateAI, an impressive new tool that is emerging out of a joint venture between Votivate, LLC (a spinoff of the Working Families Party's tech division) and a Louisiana-based AI start-up launched by a former ACORN organizer. By itself, Votivate is an alternative to VAN, the constituent-relationship-management software package that is ubiquitous on both the hard- and soft-side of Democratic and liberal/progressive political engagement. The WFP initially built Votivate because in some states, the Democratic Party would only give incumbents access to VAN. Since the upstart party often backs challengers in Democratic primaries, its candidates needed an alternative that could also access the voter file and enable all the tasks of today’s digitally-powered campaign. (Incidentally this is another reason why making third-parties electorally viable by legalizing fusion voting is so useful.) Now, as questions arise about VAN’s long-term viability since it was sold to Apax Partners, an overseas hedge fund, the WFP has been offering Votivate to more candidates. It’s one of several VAN alternatives competing for clients, alongside Open Field, Civitech and Universe.

VotivateAI is an attempt to leap-frog past that competition by delivering an innovative set of new services. Three of their offerings stood out for me. The first, Campaign Assistant, may both appeal to and threaten professional campaign managers. Essentially, if you tell it some basic information about the race you’re involved in, including whether it’s a competitive race or a below-the-radar contest, how much has been raised, and so on, the tool will produce an in-depth campaign strategy memo, which it will continue to update as news develops and variables change. Votivate’s chief operating officer, Nic Zateslo, gave me a sample of what the beta version of the tool currently produces, using some fake campaign data about a made-up candidate, “Sherry Greene,” to generate the example. Here are some snippets:

Example - Fundraising
This memo provides an analysis of Sherry Greene’s campaign fundraising strategy and performance based on the provided data. The upcoming primary election in District 3 is competitive, with a significant number of younger, online donors and older retired groups identified as potential contributors….
A. Diversification of Funding Sources: The candidate has reported raising $5,500 thus far in their campaign efforts; however, there is no information about small-donor contributions or merchandising sales within the provided data. To diversify funding sources effectively, consider implementing strategies such as reaching out to smaller donors through digital channels or creating branded merchandise for sale on the website (www.voteforcandidate.com).
II. Donor Relationships (CAMPAIGN_FUNDRAISING_GUIDELINES II) A., B.: Based on limited information provided, there is no clear evidence of long-term relationships being established between Sherry Greene's campaign team and its donors yet; however, regular communication via email and social media could help build stronger connections with supporters while keeping them informed about progress made during her candidacy journey towards winning Florida Senate seat from District 3 come August 2nd, 2024 election date mentioned above(RACE_DATA).

Example - Messaging
This memo provides a strategic plan for the campaign's messaging approach based on the provided data.
I. Identification of Target Audiences
A. Younger, online, millennial groups in District 3 concerned about poor road conditions. Demographics: age (18-35), online presence, interested in infrastructure issues. Values and motivations: desire for improved community infrastructure and transportation solutions.
B. Older, retired groups in District 3 interested in affordable housing. Demographics: age (60+), retired status, concerned about housing affordability. Values and motivations: desire for stable living situations and financial security in retirement years.
II. Core Message
A clear, concise message that addresses the needs and concerns of both target audiences while emphasizing candidate Sherry Greene's background as a high school teacher dedicated to improving her community through better infrastructure and affordable housing solutions within District 3's unique context as a predominantly republican district with specific demographic characteristics identified above (younger voters focused on roads vs older voters focused on housing). Example core message: "Sherry Greene is committed to bringing real change to our community by focusing on critical issues like improving our roads for younger generations while ensuring affordable housing options are available for all residents."
III. Key Themes - Infrastructure & Housing Solutions I) Improving Road Conditions II) Affordable Housing III) Community Development IV) Education & Experience V) Local Leadership & Representation VI) Dedication to Public Service VII) Strong Moral Foundation - Southern Baptist values aligning with constituents' values VIII) Military Background - Understanding the sacrifices made by veterans and their families IX) Educational Background - Committed to quality education X) Proven Track Record of Successful Teaching Career XI)"Vote Sherry Greene – A Voice For Our Community" or "Join Us – Together We Can Make A Difference!"

This is just a beta example working off incomplete district and race data, but it’s an indication of what’s already possible. Will campaign managers use it? They should at least try it. That’s because campaign professionals can benefit from more opinions than their own; as this recent study by David Broockman and Joshua Kalla found, political practitioners are actually pretty bad at predicting what messages will persuade the public. They found that “(1) political practitioners and laypeople both perform barely better than chance at predicting persuasive effects; (2) once accounting for laypeople’s inflated expectations about the average size of effects, practitioners do not predict meaningfully better than laypeople; (3)these results hold even for self-identified issue experts and highly experienced practitioners; and (4) practitioners’ experience, expertise, information environment, and demographics do not meaningfully explain variation in their accuracy.” So maybe getting the advice of an AI coach can only help. (This isn’t the only example of progressives experimenting with training LLMs to do political coaching--New/Mode is also in the process of developing an AI Campaign Strategist for social good and advocacy campaigns.)

Votivate AI’s second offering is a low-latency AI voice calling tool that doesn’t need a precise script, but rather can work from an overview of the campaign to have a natural-seeming conversation with a potential voter. It’s an AI campaign volunteer and unlike a human, it can make thousands of individual calls without needing a break, or pizza. I listened to a three-minute recording that Zateslo provided, and the speed and intonation of the AI agent’s banter was quite impressive. It was light-years more versatile and human-sounding than the ethereal “Ashley” voice that Civox, a London-based start-up, has created for Shamaine Daniels’ congressional campaign. (Civox says it has chosen to deliberately make its AI caller sound robotic, even though it doesn’t have to; both Civox and VotivateAI have their callers explicitly tell voters they are AI tools.)

Using generative AI to “talk” with individual voters is a significant leap past existing political AI messaging tools, like Quiller.ai, which is designed to help email writers come up with more persuasive subject lines and text, or DaisyChain, which automates text-message conversations with voters, using AI to suggest the best response (something that’s been around since IBM Watson was used to power WorkIt for the OurWalmart campaign); or SoSha (previously known as SpeechifAI), an AI writing assistant for social media posts (watch out, interns, AI is coming for your job!).

Arguably, the availability of AI campaign volunteer callers could help level the campaign playing field for underdog candidates like Shamaine Daniels, who says she’s excited about Ashley’s potential to help her reach out to more voters, understand them better, and connect in more languages (the bot speaks more than twenty). “This is going to scale fast," Ilya Mouzykantskii, the London-based CEO of Civox, told Reuters back in mid-December. "We intend to be making tens of thousands of calls a day by the end of the year and into the six digits pretty soon. This is coming for the 2024 election and it's coming in a very big way.”

While I’m intrigued by the potential of an AI coach helping campaigns build and adjust their strategies, I have very mixed feelings about an AI campaign volunteer taking over for human phone-bankers or texters. It’s true that this may give underdog campaigns a leg up. Compared to auto-dialers, which deliver pre-recorded messages, these AI callers are also far more versatile. And unlike human volunteers, an AI caller can be trusted to stay on message (assuming the tool doesn’t “hallucinate” and the recipient of the call doesn’t try to bait it into silliness.) But if campaigns start using AI this way, voters may recoil, worsening a problem that we already have thanks to all the unwanted emails, texts and calls they already get. Will campaigns use AI callers to free up human volunteers for more labor-intensive tasks like door-knocking? Or will they make less of an effort to get volunteers in the first place, since they won’t need them as much?

There’s an even more troubling question, which also arises around VotivateAI’s third offering: using AI to automatically create high-quality individualized media aimed at moving voters to action. Of course, targeting voters by specific identity characteristics is not new. But a big reason why campaigns generally refine their targeting using buckets of characteristics rather than tailoring messages to individual voters is the cost in time and money is too high. If campaigns now have the ability to create unique video messages for specific people and to do so quickly, cheaply and at scale, the potential for abuse is enormous. While campaigns may justify this approach since it can help them win elections—and who doesn’t want to find the best way to successfully reach a voter--hyper-personalized campaign messaging will give politicians who already like to pander to voters the ability to sneakily be all things to all people.

And What About the Dark Side?

The team at Votivate say they are committed to make sure their tools are only used by progressive candidates and to fight against fascism, to their credit. The founders of Civox also say they intend to mainly work with Democratic campaigns and candidates. (Unspoken are all the opportunities to make money selling these services to businesses.) Unfortunately, one of the most important lessons of the last decade also applies here. Bad actors, including foreign adversaries, will also use these tools. Which means the waters that we swim in are only going to be more polluted and unhealthy to consume.

Now, it’s worth remembering that the marginal impact of personalized persuasion in election campaigns is fairly low. The effects of most political messaging on voters fade quickly (other than deep canvassing, which is human-intensive and expensive to do well but has lasting persuasive effects). Also, something like 90% of voter decisions can be predicted simply by knowing what party they identify with; if Democrats spent one-tenth of what they spend on individual campaigns instead on party-building, improving the party’s brand identity and convincing more voters to be Democrats, they’d be in much better shape. So the chances of AI-powered hyper-personalized voter targeting “brainwashing” voters and tipping elections aren’t high.

Also, cheap and ubiquitous AI may also give voters some defenses against the worst of the coming wave. Right now people may not be able to tell the difference between a real human caller and an AI caller, but the day is probably not far off when you can install an AI assistant that will take your calls for you. Already, you can sign up for Jolly Roger, an AI-powered tool combined with a voice cloner that will take your telemarketing calls for you, and then use a series of preset expressions and topic-specific responses to keep the telemarketer on the line, wasting their time instead of yours. (I can testify from personal experience as an occasional phone-banker that some people in swing states already use such systems on their phone lines.)

Right now the main defense against AI-powered disinformation and manipulation seems to be a combination of more journalism and state regulation, neither of which seems up to the challenge. Voters are going to need to be vigilant and, hopefully, more people will share what they are seeing in their inboxes and hearing on their phone lines, which may make it harder for personalized campaigning to tip over into blatant pandering. But there’s no simple solution to the use of AI to power campaigns; the future is here and we all better watch out.

—Higher Ground Labs is working on an AI Landscape Report that will come out this spring; they’re looking for submissions of political AI tools to be included and/or examples of use cases and case studies. Here’s their December 2023 guide to selecting the right generative AI tools; do note that many of these examples are souped-up uses of statistical modeling and not necessarily as cutting edge as political coaching, voice cloning or personalized targeted videos.

—More on the “Age of AI Misinformation” from Conrad Gray of Humanity Redefined. His conclusion is the same as mine: “The advancements in AI, while remarkable, have ushered in a new wave of misinformation that threatens to undermine the very fabric of our democratic processes and societal trust. We must address this challenge with a multi-faceted strategy: embracing technological solutions for detecting fake content, fostering legislative action to hold tech companies accountable, and cultivating media literacy among the public. Vigilance, critical thinking, and a commitment to ethical AI practices must be at the forefront as we stride into a future where the lines between what’s real and what’s not become increasingly blurred.”

—New polling from the AI Policy Institute shows an emerging partisan divide over how to address the role of AI, with college-educated voters more concerned about the need for regulation.

—Unfiltered.media is offering a five-intensive training in using AI in progressive politics; apply here to participate.

—Attend: The Cooperative Impact Lab, HGL’s Progressive AI Lab and Zinc Labs are holding a half-day summit on AI in Campaigning and Organizing on January 30; register here.

—Here’s New York City Mayor Eric Adams using an AI tool to generate a Spanish-language message to constituents. I kinda miss El Bloombito.

End Times

Meanwhile, on a galaxy far, far aw.ai, an elderly couple enjoys their retirement together.

The Connector

Discussion about this post