Can Machines Catch Feelings? Simulating Social Contagion

by

Author: Denis Avetisyan

New research explores how well artificial intelligence can model the spread of emotions through online social networks.

This study assesses the limits of large language models in replicating the structural and dynamic complexities of emotion diffusion in real and simulated social graphs.

Despite growing reliance on large language models (LLMs) for simulating social dynamics, a fundamental question remains regarding their capacity to replicate the nuanced emotional landscapes of real-world interactions. This study, ‘Emotion Diffusion in Real and Simulated Social Graphs: Structural Limits of LLM-Based Social Simulation’, systematically compares emotion diffusion patterns in real Reddit discussions with those generated by LLM-driven simulations. Our analysis reveals substantial structural discrepancies, with LLM-simulated networks exhibiting limited connectivity and monotonic emotional trajectories-a stark contrast to the dense, dynamic, and heterogeneous diffusion observed in real social networks. These limitations raise critical concerns about the validity of LLM-generated data for computational social science, and suggest a need for more sophisticated simulation frameworks that capture the full complexity of human emotional exchange.

The Architecture of Feeling: Mapping Emotional Diffusion

The pervasive influence of social media platforms isn’t simply about the transmission of information, but rather the rapid and often unpredictable spread of emotional states. This propagation of feelings-from joy and excitement to anger and fear-forms the bedrock of collective behavior online, shaping public opinion, driving social movements, and even impacting real-world events. Understanding how emotions diffuse through networked communities is therefore crucial, as these emotional currents can amplify seemingly minor incidents into widespread phenomena. The speed and scale at which emotions now travel, facilitated by instant connectivity, presents a significant departure from traditional models of social influence, demanding new analytical approaches to discern patterns and predict outcomes within these emotionally-charged digital landscapes.

Initial attempts to model the spread of influence through networks, such as the Independent Cascade and Linear Threshold models, offered valuable first steps but ultimately proved limited in capturing the complexities of human interaction. These early frameworks typically treated information – or influence – as a discrete unit, propagating with fixed probabilities based on local network connections. However, they often failed to account for the crucial role of emotional states, the varying susceptibility of individuals, or the dynamic nature of social influence. The Independent Cascade model, for instance, assumed a one-shot attempt at influence, while the Linear Threshold model relied on a simple accumulation of signals. While mathematically tractable, these approaches lacked the nuance needed to reflect the rich, often irrational, ways emotions and ideas actually diffuse through populations, prompting the need for more sophisticated computational models grounded in psychological theory.

Emotional Contagion Theory proposes that feelings are, in essence, infectious, spreading through populations via mirroring and feedback loops – but demonstrating the precise boundaries of this process demands more than observation. Researchers are increasingly turning to computational models to simulate how emotions propagate through networks, allowing for controlled experiments that isolate variables like network structure, emotional intensity, and individual susceptibility. These models aren’t simply about predicting outbreaks of feeling; they aim to determine when and under what conditions emotional contagion will occur, and how it might be amplified or dampened by digital platforms. By creating virtual populations and observing the spread of simulated emotions, scientists can test the theory’s limits, refine its predictive power, and ultimately gain a deeper understanding of collective emotional dynamics – a crucial step in navigating an increasingly interconnected world.

Decoding the Signal: Methods for Quantifying Social Sentiment

Social media sentiment analysis is a critical process for identifying and quantifying public opinion regarding products, services, or events; however, the reliability of results is directly dependent on the accuracy of the classification tools employed. These tools utilize Natural Language Processing (NLP) and machine learning algorithms to categorize text as positive, negative, or neutral, and misclassification rates – stemming from factors like sarcasm, nuanced language, or contextual ambiguity – can significantly skew overall sentiment scores. Consequently, ongoing validation and refinement of these classification models, often through manual annotation of data and comparison against established benchmarks, are essential for ensuring the trustworthiness of derived insights and informed decision-making.

Lexicon-based sentiment analysis tools, such as VADER and TextBlob, utilize pre-defined dictionaries of words and associated sentiment scores to quickly assess text polarity. While computationally efficient and easily implemented, these methods often struggle with contextual understanding, sarcasm, and nuanced language. In contrast, transformer-based models like RoBERTa leverage deep learning architectures and large datasets to achieve significantly improved precision. RoBERTa, and similar models, analyze words in relation to their surrounding context, enabling a more accurate determination of sentiment even in complex sentences. Evaluations consistently demonstrate that RoBERTa surpasses lexicon-based approaches in benchmark sentiment analysis tasks, though at the cost of increased computational resources and training time.

Reddit data presents a valuable resource for gauging online emotional landscapes due to its public accessibility and diverse user base representing a wide range of demographics and interests. The platform’s structure, comprising communities focused on specific topics, allows for sentiment analysis targeted at granular subject matter, providing more nuanced insights than broad social media trends. Data extracted from Reddit comments and posts facilitates the tracking of emotional responses to current events, products, and societal issues, offering a near real-time assessment of public opinion. Furthermore, the prevalence of textual content, as opposed to image or video-centric platforms, streamlines the application of Natural Language Processing (NLP) techniques for automated sentiment scoring and trend identification.

Simulating the System: Modeling Emotional Spread in Networks

The Susceptible-Infected-Recovered (SIR) model, initially developed to model the spread of infectious diseases, provides a mathematically tractable framework for understanding emotional contagion within social networks. In this application, ‘Susceptible’ nodes represent individuals who can potentially adopt an emotion, ‘Infected’ nodes represent those currently experiencing and propagating the emotion, and ‘Recovered’ nodes represent individuals who have ceased experiencing or transmitting the emotion – effectively becoming immune to further influence for that specific emotional state. The model utilizes parameters defining the rates of emotional transmission (β) and recovery (γ) to simulate the dynamics of emotional spread across a network. By adapting the core principles of epidemiological modeling, researchers can quantitatively analyze factors influencing the reach, duration, and overall impact of emotions within a population, allowing for predictions of emotional diffusion patterns based on network topology and individual characteristics.

NetworkX, a Python package, facilitates the creation, manipulation, and analysis of complex network structures representing social relationships. This allows researchers to model emotional diffusion by defining nodes as individuals and edges as connections between them. Beyond network construction, Graph Neural Networks (GNNs) leverage the inherent relationships within these networks to predict emotional states. GNNs operate by aggregating feature information from a node’s neighbors, allowing the model to infer emotional influence and spread. The network structure-degree distribution, clustering coefficient, and path lengths-serves as input features for the GNN, enabling it to learn patterns of emotional contagion and predict how emotions will propagate through the network.

Simulations demonstrate that individual characteristics within a social network directly influence emotional contagion. Specifically, the Credibility Score, representing an individual’s influence, and the Susceptibility Coefficient, indicating their openness to emotional influence, are key determinants of diffusion rates. A random strategy for emotional transmission resulted in an average spread of 2.58 nodes within the simulated network. However, strategies incorporating theoretical principles and the eIC (emotional influence coefficient) model achieved significantly reduced spread, with averages of 1.44 and 1.12 nodes, respectively. These results highlight the importance of considering individual-level factors when modeling and predicting emotional dynamics in social systems.

The Generative Landscape: LLMs and the Future of Social Simulation

Large Language Models (LLMs) are now capable of constructing remarkably realistic synthetic social media interactions, offering a novel approach to data generation for social network analysis. Models like DeepSeek-Chat can simulate diverse conversational turns and even express nuanced emotional responses, effectively building expansive datasets without the limitations of real-world data collection. This capability circumvents challenges related to privacy, cost, and scalability often encountered when studying online social dynamics. By algorithmically creating these interactions, researchers gain the power to precisely control variables and explore specific social phenomena, such as the spread of information or the influence of key individuals, in a repeatable and customizable environment. The potential extends beyond simply augmenting existing datasets; LLMs are poised to become vital tools for proactively generating data tailored to specific research questions.

Large language models are proving instrumental in constructing expansive datasets for training and validating diffusion models designed to understand social dynamics. These models don’t simply generate text; they simulate conversational exchanges and, crucially, imbue them with emotional expression. Initial investigations reveal a pronounced positivity bias within these simulated interactions, with approximately 83.9% of replies generated from neutral prompts registering as positive in sentiment during the first round of simulated diffusion. This capacity to produce emotionally-charged data at scale offers a unique opportunity to study the propagation of feelings within virtual social environments and refine algorithms capable of interpreting complex human communication patterns.

Researchers are leveraging the controlled environment of Large Language Model-generated social interactions to dissect the mechanisms of emotion propagation, with a particular focus on phenomena like the Opinion Leader Effect. While this approach offers unprecedented ability to isolate variables, initial evaluations reveal a crucial limitation: the synthetic social networks produced lack the intricate structural complexity of real-world connections. Graph Neural Network (GNN) classifiers, designed to analyze network patterns, achieved only 75.5% accuracy and a Macro F1 Score of 0.621 when trained on these simulated graphs, suggesting that the artificial networks fail to fully replicate the nuanced relationships that govern emotional influence in authentic social settings. This discrepancy highlights the need for continued refinement of these generative models to more accurately capture the subtleties of human social structure and improve the fidelity of simulated data.

The study highlights a critical challenge in computational social science: replicating the nuanced interplay between structure and behavior within complex networks. While Large Language Models demonstrate proficiency in generating locally plausible interactions, they often fail to capture the emergent properties arising from network topology – a limitation readily apparent when modeling emotion diffusion. This echoes Barbara Liskov’s insight: “It’s one of the dangers of working with systems – you have to understand how they interact, not just how they look.” The research suggests that simply scaling LLMs isn’t sufficient; a deeper understanding of network structure and its influence on dynamic processes is essential. Good architecture is invisible until it breaks, and only then is the true cost of decisions visible.

Beyond Mimicry: The Road Ahead

The attempt to distill emotion diffusion into algorithmic form reveals, predictably, that the surface is not the structure. This work demonstrates that Large Language Models, while capable of generating locally plausible interactions, stumble when tasked with replicating the emergent properties of genuine social systems. The challenge isn’t merely to simulate feeling, but to model the constraints – the infrastructural limitations – that shape its flow. A city doesn’t become more functional by adding a new skyscraper; it requires constant, adaptive maintenance of the underlying network of roads and utilities.

Future research must shift focus from mimicking behavioral outputs to understanding the topological prerequisites for emotional contagion. Graph Neural Networks offer a promising avenue, but their current implementations often treat network structure as static. Real social graphs are constantly re-wiring, forming and dissolving connections based on subtle shifts in sentiment and trust. A truly robust simulation will necessitate dynamic graph structures, incorporating mechanisms for evolving connectivity alongside emotional states.

The pursuit of realistic social simulation isn’t about achieving perfect prediction; it’s about identifying the irreducible complexities that govern collective behavior. The goal is not to build a perfect copy, but to illuminate the fundamental principles that dictate how information – and emotion – propagates through complex networks. A more fruitful path lies in recognizing that the limitations of current models aren’t bugs, but signposts pointing towards a deeper, more nuanced understanding of social structure itself.

Original article: https://arxiv.org/pdf/2512.21138.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-26 23:08