Project INFLUENCE tests whether it is possible to influence behavior around collaborating with opposing viewpoints.

In randomized controlled trials, we treat test populations with collaboration-advocacy messaging then measure whether treated groups’ behavior differs from that of untreated control groups. We then learn and iterate.

The Goals

  1. Develop an efficiently scalable system for testing various interventions aimed at influencing partisan sentiment and behavior.
  2. Identify messaging that successfully evolves a population’s sentiment toward supporting the need to find workable joint solutions with opponents.


Phase 0.1 sentiment test

“Phase 0.1” launched April, 2022.

  • Developed initial approach of measurement system April, 2022. This system is needed to quantify attitudes before and after our messaging campaign, to assess if the campaign had any influence.
  • Launched measurement system in May.
  • Recruited resources needed to develop the messaging campaign.
  • Created messaging campaign, with five separate video ads.
  • OUTCOME: Tools tested and verified effective, messaging completed, ready to launch randomized control trial.
Sample of Phase 1 messaging.

Phase 1 – “INFLUENCE 1.0” launched May, 2023:

  • Identified test region, separated audience into Pre-test, Test and Control groups within the Meta social media environment.
  • Deployed the earlier-developed measurement system on the Pre-test group and confirmed anti-collaboration bias in region.
  • Launched messaging campaign to the Test group in April, 2023. Ran for thirty days using all five messages to reach ~25,000 subjects, roughly five treatments each.
  • Then deployed the measurement system on a random sample of 158 participants across the Test (83 participants) and Control (75) groups in May.
  • Recorded material differences in both engagement rates and sentiment between Test and Control groups.
  • OUTCOME: We viewed the differential engagement rates as encouraging and supportive of the hypothesis that the treatment is effective at improving sentiment toward collaboration. We view the differential sentiment survey results as promising, but not conclusive.
Phase 2 invite into behavior test.

Phase 2 – “INFLUENCE 2.0” conducted September, 2023:

  • Changed the goal to targeting impact on behavior (had been targeting sentiment).
  • Repeated the general outline of the Phase 1.0 test, but completely revised the post-messaging test to measure impact of the messaging campaign on changes in behavior. Also refined the messaging itself and the invitation to the post-messaging test, shortened the duration of the campaign to one week and altered the mechanism for assigning participants among the A/B groups.
  • Treated 14,000 subjects roughly 3 times each across eight days, then tested a random sample of 354 participants across Test (166 participants) and an untreated Control (188)
  • OUTCOME: Again measured meaningful differences in engagement rates between Test and Control groups, but results were ambiguous due to an error in how we assigned A/B group participants. Also failed to secure sufficient engagement with behavior test, so that data was not useful.
Sample of behavior test.

Phase 3 – “INFLUENCE 3.0 through 3.6” conducted September – October, 2023.

  • Accelerated to weekly iteration/release cycles across six weeks to rapidly develop and test methods of overcoming weaknesses identified with Phase 2.0.
  • OUTCOME: Significantly improved both click-through to the post-messaging test and submission of that test once presented, thereby dramatically increasing the number of test participants. Also fixed the A/B group assignment problem. Ultimately achieved a methodology that consistently and reliably measures an audience’s willingness to engage with pro-collaboration activities, which therefore can be used to cost-effectively measure differences in behavior between treated and control groups.

Phase 4 – “INFLUENCE 4.0/1/2” conducted October 2023 – January, 2024.

  • Utilized updated system refined in Phases 3 to once again test whether the messaging campaign has measurable impact upon behavior.
  • In two large-scale, randomized, controlled trials, treated over 106,000 subjects roughly 2.3 times each across 5-7 days, then tested a random sample of 3,335 participants across Test (1,626 participants) and an untreated Control (1,709).
  • OUTCOME: Meaningful uplift in treated subjects behavior to reduce partisan dysfunction. Subjects that had viewed the campaign were over 23% more likely than Control to both (a) consider a list of actions they can take to reduce partisanship and (b) actually take one of those actions. See the full Technical Report.

You can follow regular updates on Project INFLUENCE in our journal, “Updates and Learnings from Project INFLUENCE.

e.pluribus.US conceives, builds and tests interventions to scalably improve public attitudes toward working with political opponents.

Track the Movement

Get weekly updates on e.pluribus.US.

Project LISTEN is opening minds!

LISTEN scalably helps us understand why opponents think as they do.
Learn more >

Thoughts from across the aisle...

“The root cause of all life’s problems is looking for a simple fucking answer.”

— Anthony Bourdain

Share this page