Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Celebrity messages reduce online hate and limit its spread

Eaman Jahani, Blas Kolic, Manuel Tonneau, Hause Lin, Daniel Barkoczi, Edwin Ikhuoria, Victor Orozco, Samuel Fraiberger

TL;DR

This large field experiment shows that non-censoring, preventive messaging—in this case 42-second videos featuring Nigerian celebrities—can meaningfully reduce online ethnic hate on X in Nigeria. Using a two-stage graph-cluster randomization with hole punching, the study finds a direct decrease in hate posts of about $-2.5\%$ to $-5.5\%$ during treatment, with roughly three-quarters of this impact persisting afterward, and a notable drop in hate reposting when a larger portion of a user’s audience is exposed. Importantly, substantial indirect effects propagate through the network: higher indirect exposure to treated peers reduces hate activity among upstream non-participants, with an estimated ~53\% reduction in upstream hate reposts, indicating that amplification dynamics can be dampened even without directly targeting all users. The results suggest that scalable, celebrity-informed preventive counterspeech can complement moderation by reducing hate without removing content and can be cost-effective relative to conventional moderation approaches, though effects vary with user activity and baseline hate levels.

Abstract

Online hate spreads rapidly, yet little is known about whether preventive and scalable strategies can curb it. We conducted the largest randomized controlled trial of hate speech prevention to date: a 20-week messaging campaign on X in Nigeria targeting ethnic hate. 73,136 users who had previously engaged with hate speech were randomly assigned to receive prosocial video messages from Nigerian celebrities. The campaign reduced hate content by 2.5% to 5.5% during treatment, with about 75% of the reduction persisting over the following four months. Reaching a larger share of a user's audience reduced amplification of that user's hate posts among both treated and untreated users, cutting hate reposts by over 50% for the most exposed accounts. Scalable messaging can limit online hate without removing content.

Celebrity messages reduce online hate and limit its spread

TL;DR

This large field experiment shows that non-censoring, preventive messaging—in this case 42-second videos featuring Nigerian celebrities—can meaningfully reduce online ethnic hate on X in Nigeria. Using a two-stage graph-cluster randomization with hole punching, the study finds a direct decrease in hate posts of about to during treatment, with roughly three-quarters of this impact persisting afterward, and a notable drop in hate reposting when a larger portion of a user’s audience is exposed. Importantly, substantial indirect effects propagate through the network: higher indirect exposure to treated peers reduces hate activity among upstream non-participants, with an estimated ~53\% reduction in upstream hate reposts, indicating that amplification dynamics can be dampened even without directly targeting all users. The results suggest that scalable, celebrity-informed preventive counterspeech can complement moderation by reducing hate without removing content and can be cost-effective relative to conventional moderation approaches, though effects vary with user activity and baseline hate levels.

Abstract

Online hate spreads rapidly, yet little is known about whether preventive and scalable strategies can curb it. We conducted the largest randomized controlled trial of hate speech prevention to date: a 20-week messaging campaign on X in Nigeria targeting ethnic hate. 73,136 users who had previously engaged with hate speech were randomly assigned to receive prosocial video messages from Nigerian celebrities. The campaign reduced hate content by 2.5% to 5.5% during treatment, with about 75% of the reduction persisting over the following four months. Reaching a larger share of a user's audience reduced amplification of that user's hate posts among both treated and untreated users, cutting hate reposts by over 50% for the most exposed accounts. Scalable messaging can limit online hate without removing content.
Paper Structure (75 sections, 2 theorems, 56 equations, 22 figures, 10 tables)

This paper contains 75 sections, 2 theorems, 56 equations, 22 figures, 10 tables.

Table of Contents

  1. Preregistration.
  2. Sample selection.
  3. Treatment assignment.
  4. Outcome measures.
  5. Estimation of direct effects on participants.
  6. Estimation of indirect effects on participants.
  7. Upstream user selection.
  8. Estimation of indirect effects on upstream users.
  9. Experiment design
  10. Hate speech detection
  11. Participant selection
  12. Interaction network among participants
  13. Network clustering
  14. Cluster randomization
  15. Prosocial messaging intervention
  16. ...and 60 more sections

Key Result

Proposition 1

Under the design described in Section sec:si_treatment_randomization, the marginal propensity of each unit is where $p_t$ is the probability that a cluster is assigned to treatment and $p_{hp}$ is the probability that an individual assignment is flipped. For distinct units $i \neq j$, let $c(i)$ denote the cluster of unit $i$. The joint propensities are: For $i=j$, we recover the marginal propen

Figures (22)

  • Figure 1: Experimental sample and treatment delivery. The left panel represents the 73,136 sampled X accounts that had engaged with hate speech in blue, with other accounts in grey. Users were grouped into clusters based on pre-treatment mention ties, and clusters were randomly assigned to treatment or control. The right panel shows that between November 29, 2023 and April 17, 2024, users assigned to treatment (plus sign) were eligible to receive prosocial video ads from Nigerian celebrities directly in their feed via the X Ads platform, whereas control users (minus sign) were not.
  • Figure 2: Direct effects on hate posts and posting volume. Panels A–B report intent-to-treat percentage changes in hate posts during the campaign treatment and post-treatment periods (A) and by month (B). Panels C–D show percentage changes in posting volume. Results are plotted separately for all posts (black), original posts (green), and reposts (red). Error bars show 95% confidence intervals. Treated users reduced hate posting and overall activity during the campaign, with a large share of the reduction persisting post-treatment.
  • Figure 3: Heterogeneous treatment effects on hate posts. Treatment effects differed by pre-treatment posting volume (A), pre-treatment hate share (B), and exposure condition (C). Low and high groups for posting volume and hate share are based on median splits. Exposure combines own assignment with the share of peers assigned to treatment (see Methods). Error bars show 95% confidence intervals. Effects were concentrated among more active and less hateful users, and there is no evidence of indirect effects among participants.
  • Figure 4: Upstream users and indirect exposure. Upstream users are accounts whose posts were frequently reposted by participants before treatment. (A) An upstream user has few treated participants among their reposters. (B) An upstream user with a larger share of treated reposters, implying greater indirect exposure to the intervention. Grey squares represent upstream users; blue circles represent participants; plus and minus symbols denote treated and control status; grey circles represent non-participants; arrows represent pre-treatment repost links.
  • Figure 5: Indirect effects on upstream users. Upstream users are non-participants whose posts were frequently reposted by participants before treatment. Panels report percentage changes associated with a one-percentage-point increase in the share of an upstream user’s pre-treatment audience that was treated. Rows indicate outcome type (loss of pre-treatment hate reposters; hate reposts). Columns indicate audience type (participants; non-participants). Colors denote upstream users’ pre-treatment posting volume (all, low, high). Larger treated audience shares are associated with greater loss of hate reposters and fewer hate reposts, including among non-participants. Error bars show 95% confidence intervals.
  • ...and 17 more figures

Theorems & Definitions (2)

  • Proposition 1
  • Theorem 1: Variance and covariance under graph-cluster randomization