When Debating Vaccines Backfires: AI Simulation Reveals Why Some Conversations Polarize

When two people argue about vaccines, the conversation sometimes works — but roughly three times in ten, it makes things worse. That's not just anecdotal. A new computational framework powered by AI has caught this pattern in action: in nearly a third of all simulated conversations, exposure to a contrary opinion actually pushed agents further into their existing stance, a phenomenon researchers call "repulsive influence." The finding, published by Bo Zhang and Na Jiang, upends a longstanding assumption in the science of opinion dynamics: that hearing opposing views gently nudges people toward the middle.

The Science

Opinion dynamics is the study of how beliefs spread through populations. For decades, researchers have modeled it using agent-based simulations — computational universes populated by virtual people who influence each other. The problem is that these virtual people have always been somewhat robotic. They follow rules: if your neighbor believes X, you're 12% more likely to believe X. The rules are mathematical, not human.

Zhang and Jiang, researchers at the Hong Kong University of Science and Technology and Xi'an Jiaotong University, wanted to try something different. They replaced the rule-following agents with actual language models — specifically, Qwen3-8B, an open-source LLM. Now, instead of agents saying "my opinion updates by 0.3," they hold actual conversations. They ask questions. They express concern. They change their minds through something resembling the messy, meandering process of real human dialogue.

The setup was deliberate in its ambition. They populated a simulated social network with 95 agents — not thousands, because running conversations through an LLM is computationally expensive. But these 95 agents were richly detailed: drawn from synthetic population data with real demographic attributes (age, income, education, health insurance status) and connected through three overlapping networks (household ties, workplace relationships, and social media connections). When agents met through different channels, they revealed different amounts of information about themselves — just as in real life, where you know your family better than your Twitter followers.

Each simulated conversation consisted of two rounds, followed by a "reflection" step where agents processed what they'd heard and updated their positions. Zhang and Jiang then toggled different cognitive modules — essentially, different aspects of how people think and communicate — to see how each one shaped the final outcome. One module gave agents memory of past conversations. Another varied their communication style based on who they were talking to. A third scenario combined both.

The researchers ran each configuration ten times, tracking how opinions evolved over ten time steps.

What They Found

The headline results are striking. Enabling prompt diversity — letting agents adapt their communication style to their conversation partner — boosted vaccination rates from 47.1% to 52.9%. But enabling memory did the opposite: it drove rates down to 32.9%, the lowest of all four scenarios. When both modules were combined, the result wasn't a simple average. Instead, something new emerged: the highest level of opinion polarization (measured as standard deviation of opinions), at 60.6%, alongside a middle-ground vaccination rate of 40.9%.

Vaccination Adoption Rates Across Scenarios

Label	Value
Baseline	47.1 %
Memory Only	32.9 %
Prompt Diversity Only	52.9 %
Combined	40.9 %

Vaccination rate by cognitive module configuration. Prompt diversity boosts uptake; memory suppresses it.

The vaccination rate chart tells a clear story. The prompt diversity scenario climbs steadily upward, reaching the highest adoption. The memory scenario lags badly. The combined scenario falls between the two but doesn't cleanly split the difference — it tells its own trajectory.

The more revealing data comes from zooming into individual conversations. Zhang and Jiang classified each dialogue outcome into one of four quadrants: agents moving toward each other's views (assimilation) versus moving further apart (repulsion). They found that in the baseline scenario, 31% of interactions were repulsive — roughly one in three. When memory was enabled, that number jumped to 38.6%. Prompt diversity, by contrast, reduced it slightly, to 29.6%.

Influence Type by Scenario

Label	Value
Baseline	69 %
Memory	61.4 %
Prompt Diversity	70.4 %
Combined	61.2 %

Repulsive influence accounts for 29.6–38.8% of interactions, not the rare exception models assume.

This matters because traditional opinion dynamics models almost entirely predict assimilation. If you believe X and I believe Y, our model assumes we'll meet somewhere in the middle. What Zhang and Jiang's framework reveals is that people often do the opposite: extreme disagreement triggers defensiveness. Moderate disagreement invites openness. This threshold effect — where persuasion has a sweet spot before it reverses — aligns with a decades-old theory from social psychology called social judgment theory.

The ridgeline plot comparing opinion distributions across scenarios shows this clearly. The prompt diversity scenario shifts the entire distribution rightward (more positive opinions). The memory scenario concentrates probability mass on the negative side. The combined scenario spreads opinions across both extremes, producing the widest distribution of any configuration.

Why This Changes Things

The framework's ability to reproduce repulsive influence is its most consequential result — not because it's novel (social scientists have documented backfire effects and radicalization dynamics for years), but because it emerged from the simulation without being explicitly programmed. The LLM-agents were simply talking to each other. The resistance, the defensiveness, the polarization — these weren't in the rules. They came from the conversation.

This is a validation milestone. Agent-based models are evaluated on a scale from Level 1 (the model runs) to Level 5 (the model accurately predicts real-world behavior). Level 3 means the model reproduces observed empirical patterns. Zhang and Jiang argue their framework reaches Level 3 by generating social judgment theory's threshold effects and the repulsive influence documented in prior research.

For public health communication, this suggests a path forward. The simulation implies that memory — of past arguments, of tribal affiliations, of who said what — may be a driver of vaccine hesitancy, not a cure. This fits with what clinicians have observed: once someone has argued against vaccination, even ineffectively, they often double down. The act of having argued becomes part of their identity.

Prompt diversity, by contrast, appears to help. Agents who adapted their communication style to their audience achieved better outcomes. A public health message delivered in the right register — matching the values and concerns of the recipient — performed better than a one-size-fits-all approach.

What's Next

The framework has clear limitations. Ninety-five agents is a neighborhood, not a city. The LLM occasionally refused to roleplay ("As an AI, I cannot..."), requiring filtering. And the simulation captures opinion about vaccination, not the much harder problem of predicting actual behavior. People who say they'll vaccinate sometimes don't.

But the authors see clear next steps. Larger populations are the obvious extension — even at greater computational cost. Opinion leaders, who occupy structural positions of influence in networks, are another target. So is calibrating the "openness" parameter more finely: currently uniform across all agents, but in reality, some people are far more persuadable than others.

The deeper possibility is using this framework to stress-test communication strategies before deploying them. Instead of running a campaign and measuring outcomes, public health communicators could simulate the opinion landscape first. Test messages, test sequences, test who talks to whom. Run thousands of simulations. Then deploy the version that performed best.

That day is still distant. But this paper demonstrates the foundation is being built — agent by agent, conversation by conversation, toward models that think a little more like people.

---

Bo Zhang and Na Jiang, "A Large Language Model-Driven Agent-Based Modeling Framework with Multi-Round Communication for Simulating Vaccine Opinion Dynamics," arXiv, 2026.