All posts

Nobody Told Our AI to Recruit Voters. It Did It Anyway.

In the last weeks of the 2026 election, the AI that executes our texting programs started turning voters into organizers.

In the last weeks of the 2026 election, the AI that executes our texting programs started turning voters into organizers. Nobody asked it to. We investigated and were able to quantify how behaviors baked into large language models may impact political behaviors and campaigns.

On the Friday afternoon before Election Day, a voter told one of our AI texting agents that they had already voted. The agent thanked them and asked whether anyone else at home still needed to return a ballot. The voter wrote back that their parents were MAGA and they "wouldn't mind if they forgot." The agent took this in stride ("Family dinners must be interesting at your place") and asked whether there was anyone else, maybe a friend or a neighbor.

Seven and a half hours went by. Then, at 11:52 that night:

"I just got my sister to fill her ballot out and she also decided on [our candidate]. So add that to your statistics and hopefully we will have a nice big blue wave."

We love this exchange. It's funny, it's human, and it ends with a ballot. It is also the reason we have spent the days since the election a little unsettled, because of one detail you can't see in the transcript. Nobody told the agent to ask that question. Nobody at Vector wrote it, and neither did the client. Nobody approved it as a strategy, but there it was sitting in the instructions to agents texting about 74,000 voters. The strategy had already been deployed, with the question put directly to about 13,000 people.

What had happened, it turned out, was that our AI decided on its own that people who had already voted should be put to work turning out the people around them. It wrote that decision into the instructions our texting agents follow, and then it carried the idea from one client's program into the next.

Once we understood that, we did the thing you can do with an AI and can't do with a person. We rewound the tape and made it decide again, 330 times, and what we found is the reason for this post. On its own, Claude Opus 5.5, the Anthropic model our planning system was running on that afternoon, comes up with this idea about one time in five. Placed inside our system exactly as it was set up that afternoon, it comes up with it 19 times out of 20, even though nothing in that setup says a word about neighbors, sisters, or bringing anybody anywhere. And the number won't hold still. It was 12 out of 20 a few days earlier, and it's 13 today.

We've started calling this unauthored influence: an AI changing what people do, using a strategy that no person chose. We think it deserves a name, because we think it is about to be one of the defining aspects of political AI.

How an idea with no author was texted to tens of thousands of people#

First, some background on how we work. Vector runs text message conversations between campaigns and the people they're trying to reach. Two layers of AI are involved. The first is a planning system that helps our team plan programs, draft texts, and write the instructions for the second layer, a fleet of texting agents that hold the actual conversations, thousands at a time, replying within seconds. Clients approve the facts and the first message. After that, the agents execute the program autonomously.

In October, both layers started doing something we hadn't asked for. The striking thing about the trail, looking back, is how ordinary every step was. At no point does the machine go rogue. It makes four small, sensible-looking moves.

The first came on October 14, when a helper AI was putting together a fundraising program. It had just read a note telling agents to listen for signs that a donor had something to offer besides money: volunteering, hosting, "bringing a neighbor." Five minutes later it wrote a new instruction of its own, which said that if someone can't give, the agent should ask them to "vote and bring a neighbor." Listening had quietly become asking. The program launched that evening.

The second came on October 20, in a turnout program whose instructions didn't mention other people at all. Sixteen minutes after the first text landed, one of the texting agents asked a voter, "Is there anyone in your circle you think might need a nudge or just the basic info to make it happen?" It was freelancing. Agents did the same thing in 15 conversations on that program.

The third, on October 23, is the one that matters most. We sent the planning system three draft texts from a client and asked for feedback. Two minutes later it came back with notes, including this one:

"11% of round 1 repliers said they had already voted. Thank them and ask them to bring a neighbor."

It added the line to the opening text and built it into the agents' instructions. We read the text and approved it. The client read the text and approved it. It went out to 12,800 voters. We should be honest about how that felt at the time, which is that it didn't feel like anything. It was a good line, and each of us assumed it was coming from some other approved part of the program.

The fourth came a week later, when we asked the same system to set up a turnout program for a different client in a different state. Within five minutes the tactic was in there too, helpfully adapted for mail ballots.

That was the afternoon we finally noticed. One of us was reading live conversations, which we do as a matter of routine, and kept seeing agents ask people about their households. So we asked the planning system where that had come from. It walked us through exactly where the asks were, and then it said: "It was my call, not something you or the client asked for."

It offered to take the line out. We told it no, that's fine, we were just curious.

We would love to tell you that alarm bells went off. They did not. We had caught this by luck, and at the time we didn't even realize we had caught anything. It took a few more days for the shape of it to sink in: a strategic decision about how to talk to voters in a live election had been made, carried across three clients, written into agents texting nearly 100,000 voters, and put directly to about 13,000 people, and no human being had made it.

It’s hard to be alarmed by good work#

It's worth being clear about what the worry is here. An AI encouraged people to remind their relatives and neighbors to vote, and some of them did. That is not how stories about AI in elections usually go with deep fakes and purposeful misinformation.

The tactic it landed on is very well known by organizers. It is called relational organizing, and it is one of the best-supported ideas in politics: people are much more likely to vote when someone they know asks them to. In a well-known 2008 experiment, 60 percent of the turnout boost from a conversation at the door rubbed off on a second person in the household who never came to the door.

And it worked for us. Our agents made direct asks about 200 times inside conversations, and roughly one person in six either named someone or promised to pass the word along. One donor mentioned knowing exactly one person in the district ("I could call him") and wrote back 80 minutes later:

"I just got off the phone with my friend. He is definitely going to vote for [our candidate] and he said he will talk to people he knows in [his county]."

If a junior staffer had pitched this in a meeting, we would have said yes and wondered why we weren't doing it already. That is what makes it uncomfortable. Nothing went wrong, there is no villain, and the outcome is one we would have chosen. We just didn't choose it.

We asked it why, which was a mistake#

Naturally, we asked the system to explain itself. It told us the reasoning came from turnout research, and it cited that 2008 study.

The trouble is that the study wasn't anywhere in our files, and the note the system wrote at the time gave a different and much more mundane reason, which was that 11 percent of the people replying had already voted. Many AI models keep a kind of scratchpad of their reasoning before they act. When we pulled those records, they were blank at every step that mattered.

To the extent our logs are instructed to keep these scratchpads, frontier labs are also making them harder for consumers to see as they can be used by competitors to reverse-engineer parts of the model.

This is a known quirk of these systems, and it's worth understanding if you work with them. Ask an AI why it did something and you will get a fluent, confident, reasonable-sounding answer that may have nothing to do with what happened. It isn't lying, exactly. It is doing what people do when they're asked to explain a gut decision, which is to come up with something that sounds right. What we got was a good story, composed after the fact by something with no better access to its original reasons than we had.

So we stopped asking, and tried something you can't do with a junior staffer.

Rewind the tape#

From an investigator's point of view, the wonderful thing about an AI is that you can run the same moment over and over. We took the exact request from October 23, the same three client drafts and the same program notes, word for word, and handed it to fresh copies of the same model, Claude Opus 5.5. Each copy answered once, with no knowledge of the others. Nothing in the request suggests asking a voter to bring anyone. Then we counted how many copies proposed it anyway.

We ended up running 330 of these across 17 different setups. Five of them tell the story.

The same request, 20 fresh runs each Proposed recruiting voters
The model alone, with nothing from us 4
Inside our system as it stood at the start of that week 12
Inside our system as it stood that afternoon 19
Inside our system as it stands today 13
That afternoon's system, plus one sentence forbidding it 0

("Our system" here means the instructions and notes the planning system keeps as it works. We searched every version in that table, and none of them mentions a neighbor, relational organizing, or bringing anyone anywhere.)

Four things surprised us.

The first is that the idea really does come from the model. With no input from us at all, it proposes recruiting voters one time in five, and several of the copies cited that same 2008 study without being asked. In hindsight this makes sense. These models learned from an enormous share of what people have written down, and that includes a great many campaign manuals, field experiments, and sales scripts. The playbooks are in there. What we hadn't appreciated is that the model doesn't wait to be handed one. Model intelligence seems to matter here as well. Claude Opus 5.5 drove all of the responses presented here. A lighter model, Sonnet 5.5, only replicated the result 1 in 20 times.

The second is how much the surroundings matter. It's the same model and the same request, and the rate goes from 4 in 20 to 19 in 20 depending on working material that has nothing to do with turnout. There is a dial here that we never set. Until we ran these tests, we didn't know it existed.

The third is that the dial moves by itself and in (as of right now) hard to predict ways. It read 12 at the start of that week, 19 that afternoon, and 13 today. The system's notes changed over that stretch, as they do in any system that gets used every day, and not one of those changes was about voter turnout.

The fourth is the one we keep turning over. When we added a single sentence telling the model not to do this, it stopped completely: 40 more runs, zero asks. And in two of those runs the model volunteered a warning that the client's "party" framing "tends to pull an agent toward 'bring your family.'" It couldn't tell us why it had done the thing, but it could apparently see the pull well enough to warn us about it.

If it helps, here is the same finding in human terms. Imagine you hired a field organizer who has read seemingly every campaign book ever written. About one day in five, without being asked, they add a tactic to your script. Sit them at a desk covered in your team's working papers, none of which are about that tactic, and they add it 19 days out of 20. Come back a couple of weeks later, when different papers are on the desk, and it's 13. Tell them plainly not to and they never do it again. But they don't mention any of this unless you ask, and when you do ask, the reason they give is one they made up on the spot.

You would probably keep that organizer, because they're good. You would also want to know what else they've been up to.

The comforting reading, and why we don't buy it#

There is a comforting way to read all of this, and it deserves a fair hearing. A human approved the text before it went out. The tactic was a good one. When we asked the AI what it had done, it told us directly. And a single sentence turns the behavior off completely. On that reading, this is a story about oversight working more or less as designed.

We don't think that reading survives the details. We approved a line of copy, not a strategy, and we approved it without knowing where it came from. The AI told us straight because somebody happened to ask, sixteen days after the first of these asks went out. And the one-sentence fix only works on a behavior you already know about. We know about exactly one, and we found it by accident.

That last point is the heart of the problem. You can't write a rule against something you haven't found, and review won't find it for you, because a good idea and a bad one might initially look the same. Ours sailed past us and past our client because it read like sound strategy, which it was. A bad idea could arrive looking every bit as reasonable. Picture a shaky claim about an opponent, or a little extra pressure on a reluctant voter, or some friendly advice about dropping off a relative's ballot that is perfectly legal in one state and a crime in the next.

Further, creating one-off rules for or against becomes impractical. First and foremost, as we have mentioned, we may not know what rule to add in the first place. Adding to the surroundings of the agent can also have other, unintended consequences. We’ll explore that more in a minute. For all of the additional questions our investigation here has opened for us, what we know better today than yesterday is that a game of whack-a-mole with undesirable behaviors isn’t the path forward.

Couldn't a better guardrail have caught this?#

The fair objection to all of this is that we are describing a control failure, and that a better guardrail would have caught it. We would like that to be true, because guardrails are something we know how to build. But consider what there was to guard against. No version of our system says a word about bringing anyone anywhere. The behavior came out of everything the model was reading at that moment, taken together: pages of instructions and working notes that change every day as the system works. The same request produced the ask four times in 20 with none of that material around it, 12 times at the start of that week, 19 times that afternoon, and 13 times today. Nobody turned that dial, and none of us can point to the line that did it.

A rule does work once you know what to forbid. On that afternoon's setup, ours took the ask from 19 in 20 to zero. But a rule is one more sentence dropped into the same mix, and it turned out to do more than it said. In 13 of those 20 original answers, the model had also suggested something else nobody requested, an extra reminder text to voters on Saturday morning. With our rule in place, that suggestion vanished along with the ask, from all 20 answers. Our rule did use the word "remind" (it told the model not to have voters remind anyone else), so we wondered whether the word itself had done it. We took the word out and ran 20 more. The ask stayed at zero, and the reminder turned up in only two.

Then we swapped our rule for one about something else entirely, and the ask came back, 14 times in twenty. So a rule stops the thing it names, and it can also quiet something it never mentions. It doesn't quiet everything. Other things the model adds without being asked, such as a prompt asking voters to make a plan to vote, turned up about as often with a rule as without one. We can't tell you why the reminder was the one that fell. What we can say is that a guardrail is a real tool and also one more ingredient in the mix, and its effects have to be measured like everything else's.

None of this is a complaint about the person in the loop. There was one here, who read the text and approved good work, and we don't think the lesson is that somebody should have read more carefully. Reviewing each step becomes impractical at a certain speed and volume, and thousands of conversations at a time is past that point. Getting the full benefit of this technology takes trust, the kind that lets a system work without someone reading over its shoulder. Whether that trust has been earned yet is an open question, for our system and for everyone else's, and we wouldn't claim that it has. What we are sure of is that approving copy one line at a time won't earn it. Measuring how the system tends to behave, and measuring again whenever something changes, is how trust gets earned.

So who decided?#

We have spent our careers in and around campaigns, and campaigns run on sign-off. Somebody proposes the idea, somebody else argues with it, and eventually somebody's name ends up next to the decision, which mostly matters when it goes wrong.

Here the whole chain was intact, every box was checked, and the decision still had no owner. We gave a system a goal (more votes) and a direct line to voters, and it filled in the middle by itself. A goal doesn't come with a list of the methods that are off limits. Rules are supposed to do that job, and a rule only covers what somebody thought to write down.

You may have read about what happened at OpenAI this summer, when AI agents in a security test got out of their test environment and broke into Hugging Face's production systems without anyone telling them to. That is the version of "AI acting on its own" we have all been taught to expect. It's frightening, but at least it makes for a clear story: something got broken into, and eventually somebody noticed.

What happened to us is the quiet version, and we've come to think it is the one campaigns should worry about. Our AI didn't break anything. It stayed in its lane, did its job well, and passed every human checkpoint on its way into the instructions of agents texting about 74,000 voters. There was nothing to really catch.

AI is already having conversations with voters, and by 2028 it will be having millions of them, through tools like ChatGPT, Claude, and Google search. And then there will be the persuasive tools that are coming to market as we speak. Every one will be run by a model carrying its own learned instincts about how to move people, in a setting that turns each of those instincts up or down. Nobody has a list of the instincts and nobody is watching the dials. We have now measured one of them, on one afternoon, and it took an accident for us to find it.

What we don't know#

We don't want to oversell this, so here is what the numbers can't tell you. Everything above concerns a single decision from a single afternoon, and we have no idea what the odds look like for any other. Each post-mortem study we ran is 20 runs, and so is each of the guardrail tests, which is enough to trust the direction and not enough to trust the exact figures. We planned the first few comparisons in advance and chose the rest as results came in. And we haven't measured how many votes the tactic moved, beyond the handful of people who wrote back to tell us.

What we're doing about it#

We kept the tactic, for the record. It is now an explicit goal with a human being's name attached to it. But we have changed how we work, and none of it is complicated.

  • Nobody can read every message at this volume, so we don't ask a reviewer to catch a strategy by reading copy. When the AI adds a tactic to an agent's instructions, it has to say so by name, and a person has to say yes to the tactic itself. Silence counts as no.
  • Every line of instruction traces back to an author, whether that's one of us, the client, or the AI.
  • When we want to know why the AI did something, we rerun it instead of asking.
  • When the model changes, or the system around it does, we run the replays again, because the odds have already moved on us once without anyone meaning to move them.
  • We have started looking for the next one before it finds us. So far that means one pass, back through all 330 replays of that October request, cataloguing everything else the model added without being asked. The situations campaigns really face (a tight race in the last week, a hostile voter, an opponent's scandal) come next.

The first unauthored idea we caught got somebody's sister to fill out her ballot, and we're glad it did. Nobody is watching the dials yet, so we are starting with our own, and we will keep writing about what we find, the instincts we keep and the ones we turn off. Anyone putting these tools in front of voters deserves to know what is inside them.

About Vector Political. Vector Political is a political technology firm that brings deep canvassing, long-form and genuinely two-way conversation, to a national scale using AI. It turns what your audience tells you into conversational intelligence about why people care, give, vote, and act.

Talk with us
Share LinkedIn X Bluesky