Smith Researcher Shows Celebrity Messages Reduce Online Hate

Large‑scale experiment on X in Nigeria reveals that targeted prosocial ads reduce both hate‑speech production and its spread through the network.

Recent findings coauthored by Eaman Jahani, assistant professor in the Decision, Operations and Information Technologies Department at the University of Maryland’s Robert H. Smith School of Business, show that proactive, non‑censoring strategies can meaningfully reduce hate speech on social media—and that reaching a user’s audience may be just as important as reaching the user directly.

The results come from the largest randomized controlled trial ever conducted on hate‑speech prevention, carried out on X (formerly Twitter) in Nigeria. Jahani was part of a research team that examined whether short, prosocial video messages—delivered through the platform’s advertising system and featuring prominent Nigerian celebrities—could reduce the production and spread of ethnic hate speech. The findings show that targeted messaging can curb harmful content at scale, offering a promising complement to traditional content moderation.

Large‑scale test of preventive messaging

The researchers analyzed more than 1.7 billion public posts on X and identified 73,136 Nigerian users who had previously engaged with ethnic hate speech. These users were grouped into clusters based on their interactions, and half were randomly assigned to receive a 42‑second video encouraging empathy and restraint online.

Over the 20‑week campaign, the ads reached nearly 18,000 users and were shown more than 900,000 times. “We wanted to understand whether shifting the environment around users—not removing their content—could change behavior,” said Jahani, whose research focuses on computational social science and online networks. “Nigeria offered a particularly important setting, given the platform’s influence and the country’s history of online tensions.”

Sustained reductions in hate speech

The intervention produced measurable and lasting effects. Users who received the messages reduced their hate‑speech output by 2.5% to 5.5% during the campaign, and roughly 75% of that reduction persisted for at least four months after the ads stopped running.

The campaign also generated significant network effects. When a large share of a user’s followers saw the videos, reposts of that user’s hateful content dropped sharply—even among followers who never received the ads themselves. For the most exposed accounts, hate reposts fell by more than 50%.

Jahani said these network dynamics reveal something deeper about how social platforms shape behavior. “Researchers have mainly focused on interventions that change the behavior of the individual directly. But what if there are other incentives at play—group identity, or even monetary incentives—that make it almost impossible to change behavior directly? Luckily, we can leverage their networks to change their behavior indirectly, and this is where our study is a pioneer.”

For large producers of hate content, he noted, the intervention may not shift their behavior immediately. “But we can change the behavior of their consumers and shift their attention to non‑hate content, which in turn will influence the type of content the producer generates.”

He added that future work must also examine how these shifts interact with platform recommendation systems. “If reducing one user’s hateful content simply leads the algorithm to substitute in hateful content from another user, then it may be extremely difficult to change the overall landscape without changing the behavior of all producers—which is unlikely to be feasible. This question has profound implications for platforms’ evolution from network‑based environments to content‑dissemination systems.”

Scalable alternative to content removal

Social‑media platforms have traditionally relied on content moderation to address harmful speech, but moderation is difficult to scale across languages and contexts. Automated systems can misclassify content, and human review is costly and slow.

The study demonstrates that preventive messaging—delivered through existing ad infrastructure—can reduce harmful content without removing posts or restricting speech.

“Even outside parties like non‑profits and development agencies can achieve these goals without having direct‑to‑platform operations,” Jahani said. “It’s a proactive approach that complements, rather than replaces, traditional moderation.”

Why Nigeria matters

Nigeria has one of the world’s youngest and most digitally engaged populations, and X plays a significant role in shaping public discourse. The platform has also been a focal point for government scrutiny, including a months‑long suspension in 2021–22 and ongoing concerns about misinformation and ethnic tension.

In addition, a 2017 analysis of online hate and the government’s 2025 removal of 13.6 million social media accounts under its Online Safety Code of Practice highlights the urgency of addressing harmful content in the country’s online spaces.

Looking ahead

The researchers note that future work could test different message framings, messenger types, or AI‑driven personalization. But the core takeaway is clear: preventive, targeted messaging can reduce online hate at scale, even in high‑stakes environments.

Jahani sees both opportunity and risk as platforms evolve. “The biggest opportunity is that proactive messaging can reach users before harmful content is posted or amplified, and our study suggests it can scale through existing ad infrastructure while producing both direct and network‑level reductions in hate. AI could make this more powerful by identifying risk contexts, adapting messages to local language and culture, and testing which messengers or frames work best.”

But he emphasized that responsible design is essential. “Campaigns must be locally led, transparent, and explicitly nonpartisan, with messages co‑designed with civil society and delivered by trusted messengers from diverse ethnic, regional, and religious backgrounds. If recipients perceive the appeal as partisan or promoted by a specific group, it may lose its effect—or even lead to backlash.”

“As platforms evolve and moderation resources fluctuate, we need solutions that are both effective and scalable,” he said. “This study shows that cost‑effective and non‑invasive interventions can make a measurable difference.”

Read the paper “Celebrity messages reduce online hate and limit its spread,” via the open-access research-sharing platform arXiv.