BRENNEN

Industries / Government, education and community / Campaigns and polling

Field work

Field work is a campaign's direct contact with voters: knocking on doors, calling, mailing and texting, to persuade people or get them to vote.

7

Milestones achieved, from 2000 to 2024.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2000

    Doors beat phones

    Alan Gerber and Donald Green reported a randomized field experiment, in which voters were assigned to groups by chance, with about 30,000 registered voters in New Haven, Connecticut, before the November 1998 election. Nonpartisan reminders to vote raised turnout substantially when delivered at the door, slightly by mail, and not at all by phone. Tests like this had been rare since Harold Gosnell of the University of Chicago mailed voters in controlled trials about a century ago. By 2016 Green and Gerber had summarized about 200 such experiments; pooling 51 of them put the effect of speaking with a canvasser at 4.3 percentage points of turnout.

    doi.orgssir.org

    Credit

    Alan Gerber and Donald Green of Yale University

  2. 2008

    Telling neighbors raises turnout

    Gerber, Green and Christopher Larimer sent a series of mailings to several hundred thousand registered voters. Turnout was substantially higher among people whose letters promised to make their turnout known to their household or their neighbors.

    doi.org

    Credit

    Alan Gerber, Donald Green and Christopher Larimer, published in the American Political Science Review

  3. 2012

    France canvasses five million doors

    François Hollande's 2012 presidential campaign knocked on five million doors and chose by lot which precincts got visits, so the effect could be read in official precinct results. Vincent Pons found that the visits did not change turnout but raised Hollande's first-round share and accounted for a quarter of his second-round margin, with effects lasting into later elections. The founders of the French firm Liegey Muller Pons had worked on that campaign, and in 2016 the firm helped Emmanuel Macron's new movement go door to door with the aim of collecting 100,000 testimonies.

    aeaweb.orgeurope1.fr

    Credit

    Hollande's 2012 field campaign; Vincent Pons for the analysis, in the American Economic Review

  4. 2015

    A faked study retracted

    In December 2014 the journal Science published a study by Michael J. LaCour and Donald Green reporting that one conversation with a gay canvasser made voters lastingly more supportive of same-sex marriage. In May 2015 David Broockman and Joshua Kalla, planning a follow-up, found that the survey firm said it knew nothing of the project, and with Peter Aronow they showed the data could be reproduced almost exactly from an existing 2012 survey plus random noise. Science retracted the paper in 2015.

    LessonCheck the raw data and the vendor before building a program on a striking result.

    scunning.comdoi.org

    Credit

    David Broockman, Joshua Kalla and Peter Aronow

  5. 2016

    Deep canvassing holds up

    Broockman and Kalla ran the kind of test the faked study claimed to be. Door-to-door canvassers in Miami talked with voters about transgender people, and the voters' attitudes shifted and stayed changed for three months.

    doi.org

    Credit

    David Broockman and Joshua Kalla, published in Science

  6. 2018

    Persuasion near zero

    Kalla and Broockman pooled 40 field experiments and ran 9 of their own, and found that the average effect of campaign contact and advertising on which candidate Americans choose in general elections is zero. Effects showed up only in rare cases, such as when a candidate took an unusually unpopular position and a campaign spent heavily on finding persuadable voters. Pons's French result, which found persuasion in a first round, points the other way.

    doi.org

    Credit

    Joshua Kalla and David Broockman, published in the American Political Science Review

  7. 2024

    A deepfake robocall fined

    On 21 January 2024, two days before New Hampshire's presidential primary, 9,581 automated calls carried an AI-generated voice imitating President Joe Biden that told people not to vote, under the phone number of a local political figure who had not agreed to its use. On 8 February 2024 the US Federal Communications Commission (FCC) confirmed that its limits on calls with an "artificial or prerecorded voice" cover AI-generated voices, and on 26 September 2024 it fined the political consultant Steve Kramer $6,000,000 for the false caller ID.

    LessonExisting caller-ID and robocall law reached an AI voice call without a new statute.

    docs.fcc.gov

    Credit

    the FCC, with call tracing by the industry's traceback group

Watch for

  1. AI text bots talk back

    Firms now sell bots that take over a campaign's text-message conversation once a voter replies to a first message sent by a person. Vector Political says it sent 2.5 million texts in 2026 and had "over 20,000 to 30,000 conversations," and Convos says it worked with 10 campaigns in 2025 and about half of a 100-campaign goal in 2026. North Dakota and California require such bots to say in their first message that they are virtual assistants.

    Achieved whenan independent study measures how accurate these bots are and how their contact and turnout effects compare with human texters'.

    nhpr.org

    Claimed by

    Vector Political and Convos, to NPR, July 2026

  2. Chatbots that persuade voters

    In two papers published on 4 December 2025, in Nature and Science, researchers report that short text conversations with AI chatbots arguing for one candidate shifted the views of people who backed the other. Among more than 2,300 Americans two months before the 2024 election the shifts were a few points on a 100-point scale either way; among 1,530 Canadians and 2,118 Poles they were about 10 percentage points. The bots persuaded by piling up factual claims, not all of them accurate.

    Achieved whena field test measures the effect on actual votes or recorded turnout, since field experiments on campaign contact in general elections have found average effects near zero.

    news.cornell.edu

    Claimed by

    David Rand, Gordon Pennycook and colleagues, December 2025

Open

  1. 01

    Pool every turnout experiment

    An open, machine-readable database of every published randomized get-out-the-vote experiment worldwide, recording method (door, phone, mail, text), country, type of election, sample size, effect and its standard error, with a meta-analysis that reruns each time a study is added.

    The brief
    Credit · unclaimed
  2. 02

    Test AI practice for canvassers

    A randomized test of whether practicing with an AI "voter" makes new canvassers better at the door. Consenting volunteers are split by chance between AI practice and the usual group role-play, on a nonpartisan topic such as voter registration. Each then holds recorded mock conversations with trained actors playing residents, scored by raters who do not know which group the volunteer was in. No real voters are contacted.

    The brief
    Credit · unclaimed

Open 01 · the brief

Pool every turnout experiment

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open, machine-readable database of every published randomized get-out-the-vote experiment worldwide, recording method (door, phone, mail, text), country, type of election, sample size, effect and its standard error, with a meta-analysis that reruns each time a study is added.

Why it is still open

The pooled figures campaigns quote come from Green and Gerber's book-length review of about 200 experiments, summarized in 2016, and a European meta-analysis is cited separately. Later re-analyses pool fixed sets of these studies, but no open, updated database of the studies and their effect sizes was found, and GitHub searches for "turnout meta-analysis" and "gotv experiments" returned no projects.

ssir.org

Built with what exists

OpenAlex and Crossref to find papers; an AI model that extracts figures from long documents, such as Meta's Muse Ember (see the Note "The jobs agents repeat"), with two people checking every extracted number; R's metafor package for the meta-analysis; a public GitHub repository.

  1. First step

    Code the 51 canvassing experiments behind the 4.3-point estimate and publish the table.

  2. Done looks like

    At least 300 experiments coded from at least 10 countries, two coders agreeing on at least 95% of fields in a 50-study sample, the 4.3-point pooled canvassing estimate reproduced from the coded data, and an update every quarter.

Open 02 · the brief

Test AI practice for canvassers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A randomized test of whether practicing with an AI "voter" makes new canvassers better at the door. Consenting volunteers are split by chance between AI practice and the usual group role-play, on a nonpartisan topic such as voter registration. Each then holds recorded mock conversations with trained actors playing residents, scored by raters who do not know which group the volunteer was in. No real voters are contacted.

Why it is still open

Practice apps exist. Canvasstraining.org, listed since 2025, lets users "practice canvassing with a virtual voter" and get feedback on how well they connected. No published test of whether such practice changes how people perform was found. Experiments so far test the AI as the canvasser, not as a coach: Thomas Costello, Gordon Pennycook, Robb Willer and David Rand randomized 1,108 US adults to an AI-led deep-canvassing conversation or a matched control, in a working paper.

ai4pb.stanford.edudirectory.civictech.guide

Built with what exists

Canvasstraining.org or an open build on a current AI model with a voice; speech-to-text such as ElevenLabs Scribe v3 (see the Note "Fewer word errors"); a scoring rubric drawn from published canvassing scripts; video calls for the mock conversations; R for the analysis.

  1. First step

    Write the scoring rubric and register the design on the Open Science Framework (OSF).

  2. Done looks like

    At least 120 volunteers randomized under a design registered in advance, two blind raters per conversation with their agreement reported, and the difference between the groups published with a 95% confidence interval.