This is the 15th in this year’s series of posts by PhD students on the job market.
Our information ecology has changed tremendously over recent decades, with potentially serious consequences for ideological polarization and inter-group conflict. My grandparents used to huddle around a radio listening to the BBC World Service for updates on the Sri Lankan civil war, dependent on a broadcast from halfway around the world to learn information about their own country. Today, I can sit in the UK watching protests from Sri Lanka live streamed to my phone, inundated with opinions from any outlet or person I wish to follow. Inevitably, much of the content I process either aligns with my own views or presents the worst version of arguments from the 'other side', and I know I am not alone. Whether this phenomenon is due to psychological gravitation or algorithmic choice, we clearly need to explore the implications of listening more to views that are closer to our own.
Concerns about this pro-attitudinal bias have been growing among policymakers, but empirical evidence on implications and suggested remedies remains unclear—especially for the contexts where it matters most. We particularly lack evidence from developing countries and fragile settings—post-conflict societies, divided communities, contexts with weak state capacity—where social cohesion isn't just desirable but essential for preventing violence, or where we need broad adoption of certain practices (e.g. vaccination) to avoid large scale harms. In terms of interventions, many advocate for increasing exposure to distant viewpoints through technological regulation or other methods, while others wish to curtail or label divisive figures or contentious information. Studies restricting social media use often show modest effects on polarization, potentially because treated subjects substitute toward other confirmatory sources. Meanwhile, research on exposure to opposing views reveals that effects can vary dramatically: some inter-group exposure interventions reduce divisions while others backfire, with counter-attitudinal content actually increasing polarization.
In a world plagued by concerns about online echo-chambers, filter bubbles and outrage engagement, should policy actors deliberately expose people to diverse viewpoints or might backlash worsen divisions? How, and with whom, should they communicate on contentious issues? My job market paper provides a theoretical framework for answering these questions and highlights key considerations for network interventions in conflictual settings.
I build and compare three models that integrate pro-attitudinal attention mechanisms with network-based social learning, varying how severely people process opposing views. In the first model, people listen more to people with opinions that are closer to their own and less to those whose views are further, essentially downweighting distant views. In the second, this feature is made more extreme, with very distant views ignored entirely. In the final model, distant views are not just downweighted or ignored, but actively opposed through a backlash-style counter-reaction. This unified framework reveals which of these learning mechanisms generate different types of polarization and, crucially, which interventions might help versus harm under each scenario. I show analytically and through simulations on thousands of networks with varied and realistic features (e.g. community structure, homophily, unequal levels of influence) what factors affect long-term outcomes on the level of social cohesion. Which societies reach a consensus, fracture or even diverge to the polar extremes? And how does this differ by the learning mechanism and key features (parameters)?
The central theoretical finding is that the precise way people aggregate opinions matters enormously for polarization outcomes. Simply downweighting distant views—paying less attention to opposing opinions—still results in consensus for all connected societies. Long-term polarization only emerges when people either completely ignore views beyond a threshold (fragmenting into echo chambers that persistently disagree) or actively oppose them (generating true divergence where groups move further apart). Empirical investigation into how opposing views are processed, and where the turning points lie, therefore matters tremendously when assessing a given context’s risk of social divisions. The same network can have very different long term division outcomes depending on the learning mechanism and the range of views to which people will genuinely listen.
Figure 1. Polarization depends on the threshold beyond which backlash-style reactions are induced
Notes: The two panels show the opinion dynamics of identical networks and original opinion distributions evolving differentially due to different thresholds. In the left panel a wide range of views are received positively and the society reaches a consensus. In the right panel, the same society diverges to the polar extremes as a narrower range of opinions is tolerated.
Network structure matters too. Some networks are easier than others to break into disagreeable clusters or turn into animosity-fuelled disagreement. Since networks can be understood as a matrix capturing the level and type of interaction between each person, I use spectral methods - mathematical techniques that analyze the overall pattern of connections and how information flows through a network – to explore this. These methods allow me to create measures of network polarizability that can be computed from network data, revealing which kinds of societies are structurally vulnerable to polarization and how to identify critical "bridge" relationships holding opposing groups together. The approach provides a way to assess polarization risk and identify which relationships to protect, strengthen or attempt to create through policy intervention.
Why this matters
Social network structures play an important role in development. Societies in low and middle-income countries often exhibit clear community structures, with deep ethnic divisions and/or geographical divides and huge differences in individual levels of social influence (centrality), all of which may be reinforced or reshaped by new information sharing technologies and the way these are configured or leveraged. Consider a post-conflict society collectively navigating governance reforms or deciding on priority neighbourhoods for redevelopment. Or a situation where policymakers must communicate to skeptical communities about immunization, such as those rolling out the COVID-19 vaccination in Northern Nigeria in the wake of the 2003 Polio vaccination boycott. In both cases, well-intentioned dialogue initiatives or information campaigns could either build consensus or trigger distrust that deepens long-term divisions. My framework provides tools to help predict which outcome is likely for a given social structure and form of pro-attitudinal bias, and to design interventions accordingly.
Policy implications
When intervening in networks on contentious issues, policy makers should follow these steps:
1. Understand whether exposure to distant views causes negative reactions in the specific context and issue domain, and where the turning points lie.
The way people respond to distant views matters tremendously for the success/failure of different types of interventions on networks. Inter-group dialogue initiatives succeed or fail based on this question and this ambiguity in effects is important given the ubiquity and growing interest or these programmes. In online networks, multiple governmental organizations now seek to regulate tech companies to promote diverse exposure through algorithms while in offline networks, “across the divide” style interventions are common in conflictual settings.
Indiscriminately building inter-group connections is often a good idea and is at worst ineffective if all opinions are well-received (ineffective if they are too downweighted or likely to be ignored). But in the presence of negative reactions, this kind of intervention can be truly harmful. The wrong kind of additional connections or messaging from policy actors can interact poorly with the common homophilic patterns in developing countries to tip societies into even further divergence that is harder to come back from.
This critical first step needn’t require extensive longitudinal studies. Simple experiments or careful qualitative work exposing individuals to messages at varying opinion distances might reveal quickly whether people update toward or away from opposing views and identify critical thresholds where reactions become muted or negative – at least ruling in/out particular approaches.
2. Understand how predisposed the network is to becoming (even) more divided.
Not all networks are equally fragile. Some societies absorb substantial disagreement and still converge to shared understanding, while others fragment or diverge easily. The difference lies in structural features: how influence is distributed, how communities cluster, and critically, how many relationships span ideological divides. I show that a network's polarizability can be approximated in practice using both network topology and opinion distances. This could be done by collecting data on influence relationships and current opinions, using the ever-growing bank of existing network data or using reasonable proxies.
3. Design appropriate policy responses for the given opinion updating assumptions.
Appropriate interventions depend on whether opposing views trigger a backlash-style reaction.
When opposing views are either well-received or ignored (no backlash): Facilitating inter-group dialogue and increasing connectivity across divides generally reduces polarization. These efforts can be ineffective if too distant connections are attempted, but rarely harmful. Homophilic clustering increases risk by creating reliance on few between-group edges, so deliberately creating cross-cutting ties helps.
When opposing views trigger backlash: The same interventions can backfire catastrophically. Indiscriminately adding connections between ideologically distant groups can create hostile edges that drive divergence. My opposing distant views model shows that even a single hostile edge can transform a converging network into a diverging one when it tips the balance.
In general, who we connect matters: some relationships strengthen a networks’ resilience to polarization, others do little or can weaken it. I show theoretically that we need to focus on the building the most constructive bridge connections, rather than forcing hostile or using limited resources on ineffective ones. Operationalizing this first requires empirical work. I highlight in the paper how these people/relationships could be identified in practice.
Universal interventions that work regardless of the learning mechanism: Three strategies are shown theoretically to help across all versions of pro-attitudinal social learning, suggesting important avenues to explore and build the evidence base.
First, reducing the influence of extremist voices relative to moderates—many networks can shift from consensus to division simply by sufficiently amplifying one extreme voice. By contrast, networks with a powerful moderate voice – perhaps that of a policy or media organisation – are far harder to break, suggesting that policy actors with the resources to reach many people might benefit from adopting a more moderate framing on divisive issues.
Second, expanding the range of views that people process positively. This might be done through perspective-taking exercises, gradual exposure starting with slightly discordant views, or framing that emphasizes shared values over partisan divisions.
Third, strengthening existing or creating new agreeable bridge connections between those whose opinion distances are not so staggering that they can’t listen constructively. This is far safer than forcing dialogue between those at opposite extremes.
Key message for fragile settings: In contexts with deep divisions and backlash risk, interventions must center around strengthening and creating agreeable relationships while avoiding forcing hostile ones. Rather than indiscriminately elevating exposure between groups—which risks triggering the very divergence policy actors are trying to prevent—it’s crucial to identify and reinforce existing relationships where cross-group communication happens constructively. Using trusted moderates positioned between camps, building bridges incrementally, expanding the radius of acceptable disagreement might all be safer initial steps before attempting to span larger divides.
Nilmini Herath is a PhD candidate at the London School of Economics and the former Head of Research at J-PAL Africa. You can find her job market paper and additional materials at www.nilminiherath.com
Join the Conversation