YSocial: An artificial intelligence powered social media virtual twin

Introduction

Online social media (OSM) have revolutionized the way we exchange information. From the user’s perspective, these digital ecosystems are largely effortless (Siddiqui and Singh, 2016), enabling convenient ways of exchanging personal content (Abramski et al., 2024), seeking information, and synchronizing with others (Dávid-Barrett and Dunbar, 2012). This convenience has catalyzed a massive digital shift in social and information exchanges from offline to online settings, which has provided novel access to massive amounts of online data regarding human behavior (Stella et al., 2023). Complexity in social media strongly relates to the motto “more is different” (Anderson, 1972): the co-occurrence of many, even similar, interactions within the same context can lead to unexpected phenomena. Examples include acts as simple and seemingly insignificant as following another user or re-sharing content. Taken individually, these actions can be understood in terms of a user’s activity, psychology, and engagement (Stella et al., 2023), but when repeated by vast amounts of users, they can determine the unexpected rise and fall of a massively followed account (influencer) or can create cascades of information re-sharing. Artificial intelligence (AI) also influences online interactions since social media actions such as following, re-sharing, posting, and commenting can also be performed by non-human users. For instance, social bots can contribute positively to social media pluralism by automatically diffusing news and links external to social media platforms (Stella et al., 2019). However, they can also act by broadcasting biased content or diffusing misinformation, which exposes human users to a wider variety of harmful content (Cresci, 2020). Moreover, unlike in-person interactions, online users’ experiences are often reinforced by platform-specific algorithmic curation (Fabbri et al., 2022), that is, embedding recommender or ranking systems in social media for bolstering user engagement. These algorithms ultimately guide the attention of vast audiences toward specific trending topics, influencers, or conversations with leanings similar to one’s own (Brown et al., 2022; Cinus et al., 2022). Algorithmic curation can reinforce users’ biases while preventing healthy discussions, thus contributing to users’ fragmentation across cognitive or social groups (Morini et al., 2021). These aspects raise crucial technical and ethical questions about balancing possibilities and risks concerning the role of human–AI interactions on social dynamics regarding information diffusion, opinion formation, and mental health (Morini et al., 2024).

Why does having a virtual/digital twin of OSM platforms make sense? OSM platforms are complex systems with emergent behaviors at both structural and social levels. Data-driven studies offer valuable insights into these dynamics but face several limitations. They often rely on limited, non-longitudinal data samples, struggle to account for the “invisible” influence of algorithmic curation, and are prone to noise due to the impact of data sampling strategies. Additionally, access to OSM data is increasingly restricted, even for research purposes, despite regulatory efforts like the EU’s Digital Services Act. ¹

These challenges highlight the need for complementary data sources that offer completeness, algorithmic control, and independence from third-party access policies. To such an extent, social media digital/virtual twins (Barricelli et al., 2019; Chinesta et al., 2020) represent a promising step in this direction. Digital twins are digital replicas of physical objects, systems, or processes. Their core function is to capture real-time data from sensors, IoT devices, online platforms, and other sources in order to reflect what is happening in the physical world. Because of this continuous two-way data exchange between the real entity and its twin, they enable real-time performance monitoring, predictive maintenance, and system optimization. Virtual twins, in contrast, are models that do more than mirror an existing asset: they can also simulate alternative scenarios, supporting “what–if” analyses of conditions that may not yet exist in reality. This makes them particularly valuable for designing and testing new systems and configurations, even though the connection to a specific real-world entity may be weaker or absent. Collectively, we can therefore see twins as reference models—tools that can provide valuable insights about a physical counterpart (Schluse et al., 2018) while testing hypotheses in controlled settings: in our applicative scenario, mimic OSM dynamics by modeling those variables that open-ended data studies cannot fully capture. While not a substitute for data-driven studies, digital/virtual twins support in vitro experiments grounded in real-world data, which can enhance our understanding of social phenomena.

Why leveraging AI agents can positively affect OSM simulations? The rapid advancement of AI (Pedreschi et al., 2023) has created new gaps in understanding the latest generation of social bots, driven by large language models (LLMs) (Achiam et al., 2023). LLMs can generate convincing human texts without syntactic or grammatical errors (Achiam et al., 2023) and even reasonably impersonate specific personae (De Duro et al., 2025). Exploring LLMs’ role within social media remains a crucial yet largely unexplored research question. While research on human–LLM interactions in social media considerably overlaps with past studies about humans and social bots (Cresci, 2020), the current research literature leaves a more interesting and unique research gap. Adopting LLMs as twins’ synthetic social media agents can support constructing realistic proxies of human behavior. Milestone research suggests that LLM agent populations can exhibit emergent behaviors similar to those found in humans, such as scale-free networks, patterns of information diffusion, and affective biases (Abramski et al., 2023). Those results open new avenues for creating controlled environments where researchers can study real-world complex systems by examining which empirical phenomena emerge in twins governed by a limited set of rules.

YSocial: Where the digital world comes to life. Given the above research gap and recent AI advancements, in this work, we introduce YSocial ² , an open-source (GPL-v3), social media platform powered by LLMs. In its current release (v3.0, 12/2025—available for macOS, GNU/Linux, and Windows), YSocial implements a virtual twin, with limited ingestion of real-world data, focused on providing a realistic social simulation framework—while maintaining the potential to evolve into a full digital twin.

In the subsequent sections, we first review the research literature that has been pivotal in developing YSocial. Then we present a technical overview of the system and illustrate its capabilities through a case study to provide better context for its behaviors. Subsequently, we explore the potential impacts of an OSM twin on computational social science and conclude with a brief discussion of the project’s future evolution.

Related works

This section provides an overview of the relevant literature concerning the development of YSocial. Focusing on simulation studies (Pappalardo et al., 2024), we delve into agent-based simulations and discuss recent advancements since the advent of LLMs.

Computational social science. The field of computational social science (Conte et al., 2012)—at the crossroads of complex systems, network science, and big data analytics—seeks to discover universal laws of human behavior through computational methods. In the natural sciences, experimental designs allow stringent control over variables, enabling precise isolation and examination of causal relationships. This control aids in replicating experiments and validating results. On the contrary, in the social sciences, experiments are challenging due to both practical and ethical constraints. Fortunately, the advent of the digital age has not only led to the production of an unprecedented amount of data traces from human activities, but enhancements in computational means have also opened possibilities for more complex experimental setups, overcoming previous difficulties.

Coarsely defined, social simulations (Squazzoni et al., 2014) are computational models designed to replicate the behavior and interactions of individuals or groups within a social system. Simulations can be implemented using different frameworks, agent-based modeling (ABM) (Bianchi and Squazzoni, 2015) being one of the most commonly employed. In ABM, humans—and other individual elements, such as institutions, technologies, etc.—can be seen as software entities called agents, which can be part of an environment and can be embedded in an interaction space, influencing their decisions and behaviors. As the latter is often represented as a graph—being fully connected (Sîrbu et al., 2019) or having a more complex configuration (Pansanella et al., 2022)—agents can have a set of edges, representing their relationships, which can be static or dynamics and may hold attributes representing specific characteristics, for example, strength, trust, or influence. For example, it may be that the network of interactions evolves as a simulation progresses as a consequence of a pre-defined rule of the model, as in the Barabasi–Albert model (Barabási and Bonabeau, 2003) of network formation, or of the evolution of other agents’ attributes, as in adaptive networks models (Pansanella et al., 2022). Agents can, in fact, be enriched with attributes, such as demographic factors or cognitive biases (Deffuant et al., 2000), which are mainly static during a single simulation and with states, that is, dynamic attributes dependent on social influence, representing factors such as infection (Lorig et al., 2021), adoption (Serrano and Iglesias, 2016), opinions (Deffuant et al., 2000), or emotions (Schweitzer and Garcia, 2010), generally modeled as simple numerical variables and often the core element of the social simulation study. To execute the social simulation, each agent follows a set of rules—algorithms dictating when an agent becomes active, the actions they can take, and the impact of these actions on themselves and the population—which can be more or less strict, giving the agent different degrees of autonomy. Social simulations have been used to study epidemics (Lorig et al., 2021), diffusion of information (Guille et al., 2013), belief and emotion dynamics (Deffuant et al., 2000), economics and finance (Axtell and Farmer, 2025), urban dynamics (Cornacchia and Pappalardo, 2021), political sciences (Epstein, 2002), and other phenomena.

Simulations of online platforms. In recent years, a great deal of attention has been dedicated to studying such social dynamics within socio-technical systems (Vespignani, 2012). Many simulation studies try to explain online-specific phenomena, such as polarization, echo chambers, or misinformation diffusion, without incorporating any specific characteristic of online environments (Peralta et al., 2024). Online platforms are often explicitly modeled by incorporating recommender systems into the simulation. In several studies, a recommender system is used as a parameter in simulations to modify the probability of agent interactions (Cinus et al., 2022; Pansanella et al., 2023), to alter the ordering of news feed posts (Gausen et al., 2022), or to create new links in the network (Chitra and Musco, 2020). Other works simulate online social platforms by incorporating specific behaviors such as posting, retweeting, and user sessions, often tailored to specific platforms (Nasrinpour and Friesen, 2016). Simulations also use data from online social platforms to calibrate micro-behaviors and/or validate macro-results (Monti et al., 2020).

AI-enhanced social simulations. Recent advancements have introduced AI agents that leverage generative models—especially language models (Brown et al., 2020)—to simulate realistic human behavior, achieving believable individual and collective behaviors (Park et al., 2023). Among these, LLMs have garnered significant attention for their ability to produce emergent behaviors akin to human societies (Leng and Yuan, 2023). This framework is in its infancy; thus, we see very different methodologies and implementations that make it difficult to compare the results obtained in the same way as in classical agent-based models. Tested models vary from closed-source, such as the GPT series (Mou et al., 2024), to open-sources, such as Llama, Mistral, and others (Breum et al., 2024), or a mixture of the two (Shu et al., 2024). In general, the aim of these latter studies is to compare results across different models. Agents are often assigned different “personas”—name, age, gender, interests, personality traits, and political leanings—through textual prompts. Environments can be absent (Xie et al., 2024), with just a population of agents interacting with humans or each other, but also very complex. For example, in Park et al. (2023), the authors created Smallville, a sandbox environment resembling a small village where agents are represented similarly to characters of The Sims. In Mou et al. (2024), there is a Twitter-like environment with timelines where LLM-agent-generated content is published and can be read by other agents. In Leng and Yuan (2023), the authors examined whether LLMs can display social learning, preference, and indirect reciprocity. Recent studies also focus on understanding the interplay between social dynamics and language (Breum et al., 2024), for example, toxicity levels in online discussions (Tornberg et al., 2023), which was unfeasibly exploiting classical agents. Generative agents show human-like collective behaviors such as forming scale-free networks (De C Gatti et al., 2013) and information diffusion (Gao et al., 2023). In the absence of possible comparisons with classical agents, generative agents showed the ability to generate persuasive arguments, in line with psycho-linguistic theories of opinion change (Breum et al., 2024; Monti et al., 2022), and less toxic behaviors, being more polite and respectful than real social media users.

Despite having advanced social simulations with generative agents, recent works are still far from creating a social platform twin. Digital/virtual twins are already vastly employed in different fields, from engineering (Tao et al., 2018) to healthcare (Sun et al., 2023) and Smart Cities (Grubel et al., 2022). Having a social platform twin would allow researchers to perform different social experiments, investigate feedback loops between social and technical components of such systems, test different what–if scenarios, and use insights to inform policymakers in a more cost-effective and ethically compliant way. However, creating such a twin is not a trivial task for the additional complexity of the system society. LLM-enhanced agents have been extensively explored in their ability to autonomously replicate human characteristics when prompted and/or fine-tuned. However, to date, there has been no social platform twins mirroring characteristics of one or more platforms and enriched with agents that mimic human behavior—which is the core of our proposal.

YSocial—virtual twin

YSocial is meant to allow for profound flexibility in simulation scenario development.

To fulfill such a goal, we designed our social media platform twin to be composed of four interacting modules:

a REST API server designed to expose all those primitives describing the actions implemented by the social platform and to store the simulation data;

Given its modular design, YSocial allows the deployment of multiple clients—each handling its own agents’ population—on each simulation, thus distributing the computational load needed to simulate agents’ behaviors and supporting scalability. Moreover, it is worth underlining that YSocial makes it possible to leverage both commercial LLMs (i.e. OpenAI) and self-hosted ones ³ (e.g. served through Ollama, LM Studio, or similar services).

In the following, we discuss the REST API server (i.e. y_server), the simulation client (i.e. y_client), and the zero code web interface modules, describing their rationale and the implementation choices performed.

y_client: Simulating social agents

As discussed, the primary function of y_server is to expose those primitives that LLM agents can use to implement their social actions. If y_server can be seen as a “light” interface toward the database storing the simulation data, the y_client is the module that implements most of the business logic involving LLM interactions. In particular, it controls the agents and their characteristics, implements the agents’ actions selection process (i.e. the strategy used to select which endpoint exposed by y_server to connect to and how to interpret the obtained responses), and defines the simulation loop for the handled agent population. In the following, we review the main facilities offered for each task—starting from the most important component of a social media digital twin: the agents that populate it.

LLM-powered agents

YSocial’s main peculiarity lies in the synthetic agents that populate its simulations. Conversely, from a classic mechanistic simulation environment, where action selection is often implemented as a sampling from a given probability distribution—that implicitly fixes their expected volume distribution—YSocial leverages LLMs as chaos engines, delegating to each agent the choice on the actions to take when activated. Moreover, to increase the heterogeneity of agents’ behaviors—and generated contents—it leverages the ability of LLMs to role-playing different users’ profiles (i.e. persona).

Agent profile. While instantiating an agent, y_client allows to specify the following dimensions:

- Agent name and ownership: Name is the unique identifier for the generated agent, while ownership links the agent to the client execution that generated it. This latter information is useful for checking the simulation status while multiple clients (having different configurations) are active.

- LLM model: This field identifies the model for interpreting the agent’s actions. A different model can role-play each agent.

- Age, spoken language, education level, nationality: Demographic information has been shown to affect the content generated by LLMs. In particular, age and education can be used to tune the expected writing styles of agents.

- Political leaning, toxicity level, and activity level: These dimensions are crucial for social simulations to: (i) generate social interactions centered on sociopolitical themes, (ii) have a predefined level of contents’ toxicity, and (iii) integrate agents with heterogeneous posting/interacting behaviors.

- Topic(s) of interest: The list of interest topics represents the background of the agent and acts as a starting point for the generation of posts and to make consistent contributions to discussion threads. Specifying a reasonable set of topics per agent increases the heterogeneity of produced texts. Agents’ interests evolve as a consequence of their actions.

- Recommender system(s): As previously described, each agent is assigned the content/follow recommendation algorithms that regulate its platform experience. Curation strategies can be updated during the simulation.

- Psychological traits: The big five personality traits, also known as the five-factor model (FFM) (McCrae and John, 1992) describe five broad dimensions of human personality. FFM is used to fine-tune the agent’s expected behavior while interacting with other users’ contents (and generating its own).

Agents’ profiles concur with the description of detailed and consistent traits that the LLM can interpret in its role-playing. It is important to underline that the simulated agents do not have a specific goal other than participating in social dynamics. However, in an online social environment, some users have their own agenda (promoting specific content, doing propaganda, etc.), an example of such a behavior being news pages.

OPEN IN VIEWER

Pseudocode of the Simulation Loop.

Anchoring simulations to reality: News pages. In a social media environment, news pages provide elements to linchpin discussions among users on real-world events and, in doing so, introduce novel discussion topics. In YSocial, news pages are agents that, conversely from regular ones, pursue a given objective: spread content generated by real-world outlets. To fulfill such a goal, each page is linked to a really simple syndication (RSS) feed of a specific outlet (e.g. CNN and Fox News). Leveraging such information sources at a predefined rate, each page shares real-world news and comments as the outlet’s social media manager would. As standard agents, pages can also have a specified political leaning to present news to their target audience more realistically.

Other agents can share and comment on page-produced content, thus replicating information cascading processes. Moreover, RSS feeds are also used to gather contextualized images that YSocial agents can use to generate multimodal posts.

Multimodal post contents: Sharing and commenting images. YSocial agents, when asked to write a post, receive as input their profile and either (i) a set of topics they are (currently) interested in, or (ii) an image retrieved from the meta fields of an RSS feed. In this latter scenario, the image URL is passed as input to a dedicated LLM model (minicpm-v) (Yao et al., 2024) that generates a textual description of the image. Such a description is then passed to the LLM model, impersonating the agent to allow him to generate a post and thus start a discussion on it. Such a choice has been made to make multimodality available independently of the capability of the chosen generative AI model.

Agents’ interests, dynamics and generated content annotation. y_client annotates agents’ generated contents to refine the simulation and provide the end users with standard indicators to enrich their analyses. In particular, for each post/comment, y_client computes:

The VADER (Valence Aware Dictionary and sEntiment Reasoner) sentiment scores (Hutto and Gilbert, 2014);

The perceived toxicity (Google’s Perspective API) (Lees et al., 2022);

The elicited emotions (GoEmotions taxonomy) (Demszky et al., 2020);

(In case of RSS gathered news) the topics expressed by the article. ⁴

In YSocial, generated posts, articles, and shared images are associated with specific topics: posts inherit the topics used in the LLM prompts, articles have topics identified through topic modeling, and shared images adopt the topics of their related articles. This enables YSocial to track agent–topic interactions throughout the simulation. Topic annotation plays a key role in updating each agent’s interests by recording the topics they interact with, the frequency of these interactions, and their timing.

On the one hand, news pages introduce new topics to the discussion. Conversely, agents’ interactions build dynamic profiles that leverage their activity history. This history, subject to a forgetting window to simulate fading interests over time, is used to sample topics for generating new content. Agents select topics in proportion to their recent interactions, increasing the consistency of their behavior. To further maintain coherence, YSocial combines selected topics with the agents’ recent sentiments toward them when crafting new posts—either computed via VADER or inferred from the agents’ reactions to content on those topics, ensuring the generated content aligns with the agents’ established attitudes.

Simulation loop

Once the rationale behind the agents’ characterization and the rules that govern their behavior is defined, we discuss how (each instance of) y_client orchestrates the agents’ activities. To such an extent, we review the configuration variables exposed by y_client and briefly discuss how the simulation loop unfolds. Simulation workflow. Algorithm 1 reports a prototypical simulation workflow implemented by y_client. The simulation starts by generating and registering the agents (and pages) to the y_server. 5 After the initialization phase, the simulation loop unfolds by activating a percentage of agents coherent with what is specified in the configuration file for each slot. Before each round of activity, the y_client instance synchronizes with the y_server simulation clock to avoid issues in distributed client settings. At the end of each day, the agent population is increased according to what is specified by the new_agents_per_iteration simulation parameter. Symmetrically, agents churn are evaluated to allow agents to leave the simulation.

Such a pipeline—that can be extended and refined to better match specific needs—along with agent populations segmentation over multiple y_clients allows the description of complex experimental settings addressing topic-specific simulations.

Zero-code experimental platform

The primary goal of YSocial is to serve as a tool for supporting computational social science research. Given the multidisciplinary nature of such a research field, an important dimension to consider while developing a twin lies in implementing strategies to lower those technological barriers that might hinder non-technical researchers from using it. To such an extent, YSocial integrates a fundamental component: the zero-code platform, consisting of two modules: the admin panel and the simulation interface.

Admin panel. To support experiment configuration, YSocial includes a flexible web interface (see Figure 1) that allows researchers to:

Manage user accounts, including both super users and regular users—the former with experiment creation capabilities, the latter being only authorized to take part in, specific, hybrid simulations.

Create, edit, delete, import, and export agent populations.

Create individual agents.

Create news pages.

Create, configure, run, and import/export experiments.

OPEN IN VIEWER

Figure 1.

YSocial admin panel. (a) Dashboard to control experiment execution and manage deployed large language models (LLMs). (b) Ad-hoc agent creation form. (c) News page details with content previews. (d) y_client details and advanced configuration panel.

To each experiment, the platform associates a y_server instance—implemented as an independent application that can be spawned and terminated when needed. Populations abstract agents’ segmentation: each one of them can be configured to follow specific characteristics (demographic, agents’ temporal activity patterns, LLM server to use, recommendation systems deployed, LLM prompts to adopt, etc.) and can be associated with multiple experiments. News pages and ad-hoc agents can be deployed to multiple populations simultaneously. Moreover, ad-hoc agent creation allows the description of very specific agents (i.e. politicians and journalists) by providing a dedicated pre-prompt to override the standard persona based one. Once a population is associated with an experiment, a y_client is created. Such a client can be fine-tuned by specifying additional information—such as the initial social network connecting the agents it controls—and, once executed, will follow the simulation loop described in Algorithm 1. Experiments’ simulations can be run, stopped, and resumed from the admin dashboard. Finally, the admin panel allows the loading of the simulation interface (for running/stopping experiments) to allow inspection and hybrid human–agent interactions.

Simulation interface. YSocial offers an X-like interface, as shown in Figure 2. This interface allows users to interact with a YSocial simulation by performing the same actions available to the agents. Additionally, to support exploratory analysis of the generated content, all post and comment annotations—such as elicited emotions, identified topics, and their associated sentiment—are visually integrated into the interface. These annotations include hyperlinks, enabling users to navigate to other content sharing the same metadata. Moreover, articles shared by news pages and images are embedded within post elements, integrating a direct link to the original source.

OPEN IN VIEWER

Figure 2.

YSocial simulation platform: (a) Agent timeline with a recommender system-curated feed and contact suggestions. (b) Agent’s profile page displays demographic information, generated content, and interactions. (c) Agent’s follower–followee details with individual activity summaries. (d) News page profile.

Case study: Political debate arena

To briefly showcase the data a YSocial simulation makes available, we describe a simple case study involving a relatively small number of LLM agents debating general politics-related themes.

Generated data

To provide compelling insights on the properties of simulated data, we now focus on three analytical dimensions: (i) global characterization of observed trends, (ii) comparison of Democrat/Republican generated contents, and (iii) generated social structure.

Global characterization. As a first step, we analyze key distributions of the agents’ generated data, focusing on their patterns and the emerging trends that arise organically and are not hard-coded into the simulation.

Figure 3(a) shows the LLM agents’ hourly activity rates as fitted by a data-driven observation made on the full dump of a similar service, BlueSky (Failla and Rossetti, 2024). That choice was made to better simulate agents’ activities’ circadian cycles and realistically modulate the expected number of online users.

OPEN IN VIEWER

Figure 3.

Simulation statistics. Top row: Percentage of active users at the hourly level. Middle row: Cumulative decreasing probability of the number of posts per user, number of comments per user, number of hashtags per user, number of mentions received by users, and distribution of discussion thread size. Bottom row: (c) Distribution of the number of comments per post and (d) the number of recommendations per post.

Figure 3(b) shows the cumulative decreasing distributions of content types (posts, comments, hashtags, mentions, and threads) in agents’ activities. Plotted on a double logarithmic scale, all distributions span multiple orders of magnitude, revealing rare elements with extreme values. Hashtags and mention frequencies highlight trending topics and a right-skewed probability of tagging, reflecting influencer-driven engagement patterns. The long-tailed distribution of posts per user indicates a few users drive most discussions, while the distributions of comments per post and thread lengths suggest the presence of viral content, where rare posts trigger widespread interactions.

Finally, Figure 3(c) and (d) delves deeper into the analysis of viral content, focusing on the distributions of comments and recommendations per generated item (post or comment). These distributions reveal a highly heterogeneous pattern of interactions, aligning perfectly with the expected outcomes of the imposed recommendation system. The ReverseChronoFollowersPopularity recommender produces a skewed distribution of post and comment visibility, increasing the likelihood that certain content (i.e. the ones popular among the agent’s contacts) gains more attention, sparking comments, shares, and reactions.

Populations’ generated contents. The simulation was designed to observe interactions between the Democrat and Republican populations. Figures 4 and 5 compare how the two populations engage with news pages’ content, focusing on the emotions elicited, topics discussed, and hashtags used in their posts and comments.

OPEN IN VIEWER

Figure 4.

Comparisons between Democratic (blue) and Republican (red) agents. Top row: Interactions ((a) comments and (b) reactions) with news page contents. Bottom row: (c, d) Most frequent emotions elicited in generated textual content.

OPEN IN VIEWER

Figure 5.

Comparisons between Democratic (blue) and Republican (red) agents. Top row (a, b): Most common agent interests. Bottom row (c, d): Most frequently used hashtags.

Figure 4(a) and (b) unveils an unbalanced information diet, with each population predominantly commenting on/reacting to content published by their preferred news outlet. Particularly, a notable pattern emerges when considering agents’ reactions to the content promoted by these pages: each population tends to support the narrative of its affiliated broadcaster while reacting negatively to content from the opposing side. Figure 4(c) and (d) depicts two populations that exhibit only slight differences in the overall ranking of their top-10 most frequently elicited emotions, with no clear patterns distinguishing the content based on political leaning.

Similarly, Figure 5(a) and (b) shows that the top 10 discussed topics remain consistent across both populations, suggesting that the content published by news pages sparks conversations that transcend political boundaries. However, the hashtags generated by the agents, shown in Figure 5(c) and (d), reveal a stronger connection to political leanings. While agents from both sides discuss similar topics and express comparable emotions, they contextualize their contributions with hashtags that align closely with their political beliefs, reinforcing their stances even when interacting with news pages’ content. For example, when discussing economy, Democratic agents use hashtags such as #RenewableEnergy and #GoGreen, which are entirely absent from the Republican discourse. Conversely, Republican agents, reflecting their political perspective, employ election-related slogans such as #AmericaFirst and #Trump2024.

Social structure. A key aspect of YSocial simulations lies in the topological patterns generated by agents’ interactions. As previously discussed, our experiment was seeded with two distinct LFR networks representing segregated populations.

The two recommendation systems influence topological evolution differently: the follow (CommonNeighbors) model strengthens existing communities by reinforcing connections among agents with mutual links, while the content model fosters cross-population interactions by allowing agents to follow or unfollow peers based on content engagement. The former densifies ideologically homogeneous groups, whereas the latter promotes content diffusion and the emergence of a single connected component.

Figure 6 illustrates the social topology shaped by the two recommenders at the end of the 60-day simulation. As expected, the two agent populations merged, generating a cohesive network topology while maintaining a certain degree of polarization—and a scale-free degree distribution. The network statistics—reported in the table for four different stages of the simulation at 15 (A), 30 (B), 45 (C), and 60 days (D)—reveal a consistent pattern in the system’s growth. The social graph’s density and diameter remain almost constant, even considering agents joining and leaving the system daily. Degree assortativity tends to maintain neutral to slightly negative values. Interestingly, label assortativity (computed on political leaning) decreases as the simulation progresses, indicating a global increase in heterophilic mixing. This trend aligns with expectations given the initial topological conditions. However, the populations remain predominantly clustered by political leaning, reaching a final assortativity score of 0.52 .

OPEN IN VIEWER

Figure 6.

(a) Social graph visualization: colors identify agents’ political leanings (red Republican, blue Democrat). (b) Network statistics for consecutive observation slots (from 15 days to the simulation conclusion). | N | number of nodes, | E | number of edges, A degree assortativity, L A political leaning assortativity, and d diameter.

Discussion

In this section, we discuss the potential impact of digital/virtual twins of OSM, such as YSocial, across various research fields—namely Network Science, Social AI, Natural Language Processing, and Human and Machine Psychology.

Network Science is an interdisciplinary domain that focuses on understanding and analyzing complex networks’ structure, dynamics, and function. Online social networks have been the research focus in several directions, such as influence and information spread, community detection and analysis, homophily, polarization, etc. A platform such as YSocial offers a remarkable tool for designing controlled experiments and serves as a null model for targeted studies. The network structure is shaped by social mechanisms such as the preferential attachment (i.e. rich-get-richer) and triadic closure (i.e. friends of my friends are my friends), observed in real networks (Cheng et al., 2019; Huang et al., 2014). However, studying real social networks is often limited to incomplete snapshots, with little insight into external factors, such as meeting people outside the network. Recent studies (Papachristou and Yuan, 2024) show that LLM-powered agents can replicate these dynamics in simplified scenarios focused on topology.

YSocial offers the opportunity to explore how different parameters influence the emergence of such phenomena. Another key area is information diffusion, which is crucial for understanding topics such as fake news spread, social contagion, and influencer identification.

Social AI studies human behavior in digital environments and the interplay between users and algorithms. Research has shown that online platforms shape individual behavior, collective decision-making, and societal outcomes (Nguyen, 2020; Pedreschi et al., 2023). A key focus is online information environments, where digital spaces can amplify biases, fueling misinformation, polarization, and echo chambers (Cinelli et al., 2021). Yet, the same dynamics can foster positive outcomes, such as supportive communities (McPherson et al., 2001; Newman, 2003). Central to this field is human–AI co-evolution, which explores feedback loops between users and algorithms (Pedreschi et al., 2023). While observational and controlled studies offer insights, they face causality and data access limitations. Simulations help fill this gap, but often simplify real-world complexity (Pappalardo et al., 2024).

In this context, YSocial provides a controlled environment to study user–algorithm interactions and test interventions. It allows researchers to model diverse scenarios—from social clustering to content exposure—and quantify their impacts on polarization, echo chambers, and misinformation. YSocial also supports integrating real-world data, enabling more realistic simulations of phenomena such as news spread and moderation strategies.

Natural language processing (NLP) and content analysis. NLP has long grappled with challenges such as data scarcity and sparsity, particularly in domain-specific or multilingual tasks. For example, hate speech detection (Kovács et al., 2021) suffers from limited data for recognizing nuanced linguistic cues. While data augmentation methods offer partial solutions (Chen et al., 2023), some tasks require insights beyond textual features, benefiting from interdisciplinary approaches. Stance detection exemplifies this, as it often relies on social media data to infer user positions on controversies (Küçük and Can, 2020; Mohammad et al., 2017).

YSocial could address two major issues: generating synthetic datasets to supplement scarce annotated data (Hardalov et al., 2022) and integrating network features with language cues to improve detection. Similar benefits extend to argumentation mining (Lippi and Torroni, 2016), which examines argumentative structures in the text. Another key application is sentiment and emotion analysis, where transformer-based models have shown promise (Krugmann and Hartmann, 2024). YSocial facilitates this by generating annotated text on the fly, simplifying the process. Additional NLP tasks such as named entity recognition (Li et al., 2020) could further enrich the synthetic datasets. In natural language generation, YSocial provides a playground to test LLMs’ abilities in social contexts, such as maintaining coherence across long discussions or showcasing creativity and reasoning in responses. This opens new avenues for understanding LLM behavior in dynamic, multi-agent environments.

Human and machine psychology. Psychology studies human mental states and behavior, while machine psychology investigates AI through cognitive and psychological frameworks (Abramski et al., 2023).

YSocial offers a unique environment for exploring human and machine psychological patterns. Social media significantly impacts emotions: positive interactions boost mood, while negative ones, such as cyberbullying, can cause distress (Valkenburg et al., 2006). Online disinhibition, where users behave more aggressively due to perceived anonymity (Suler, 2004), could be studied with YSocial by simulating topic-driven conversations, identifying behaviors such as trolling or offensive language (Moor et al., 2010). Another key phenomenon is social comparison. Festinger’s theory suggests people evaluate themselves relative to others (Festinger, 1954), a process amplified on social media, often leading to negative self-assessments (Vogel et al., 2014). YSocial could simulate scenarios where LLMs undergo social comparisons, helping researchers identify protective interventions that mitigate negative effects—insights that could be applied to safeguard human users from harmful comparisons. Finally, YSocial could aid in combating misinformation, which often spreads due to cognitive overload, not just lack of understanding (Pennycook and Rand, 2021). LLMs, less prone to these limitations, could simulate environments to test how belief formation affects misinformation spread. Investigating when LLMs fail to detect misinformation would provide crucial insights for designing AI-assisted tools to reduce exposure to false information. Despite lacking human cognitive biases, LLMs might still amplify misinformation due to limited reasoning capabilities, a critical area for further research.

Conclusion and future works

In this work, we introduced YSocial, an LLM-powered virtual twin of OSM platforms. YSocial enables the design and execution of complex agent-based social simulations, avoiding mechanistic or deterministic solutions while allowing researchers and practitioners to conduct controlled what–if scenario testing. To address the limitations of existing solutions, YSocial integrates social interactions and content generation into a unified process, ensuring these dimensions naturally co-evolve during data generation. To complement our discussion of YSocial’s platform design, we presented a case study simulating social media dynamics in a two-party political debate. Our analysis highlighted the platform’s ability to replicate emergent social patterns influenced by AI agent characterization and recommender system curation, further supporting its application for studying diverse social scenarios.

YSocial represents a first step toward defining a mature tool to support computational social science studies. As its current stage YSocial can be classified, due to its limited data ingestion from the real world (restricted to RSS news feeds), as a virtual twin. However, we plan to gradually move toward a digital twin by strengthening the link toward online platforms (i.e. X and Reddit), integrating a more structural data ingestion pipeline to better anchor simulations to real settings. Considering AI agents’ sharp evolution and complexity, validating a social media twin and assessing its limits in replicating human behaviors are crucial tasks to be addressed (Taillandier et al., 2025; Wu et al., 2025). To such an extent, as future work, we plan to apply YSocial to multiple multidisciplinary research endeavors to empirically validate our platform across various use cases with domain experts’ support. Additionally, we plan to enhance YSocial by introducing new simulation layers and features, such as goal-oriented agents (e.g. moderators, bots, and gatekeepers), expanding agent actions and consistency (e.g. integrating individual temporal-evolving knowledge graph representation of the world), and enabling users to select between different platforms—for example, integrating forum-like dynamics and developing full multimodal simulation environments allowing image/video generation and sharing. Finally, our work conceptually aligns with the aims of broader social mining frameworks, that is, PNS (Planetary Nervous System) (Giannotti et al., 2012), where privacy and ethical safeguards are central to large-scale data analysis. In this context, the role of prompting in YSocial is critical. While the current version deliberately relies on generic prompts to simulate general personae and avoid privacy risks, we acknowledge that more specific prompts (e.g. imitating real individuals or highly detailed profiles) could raise serious privacy concerns—especially while moving toward a full digital twin. Exploring how YSocial might handle such scenarios, possibly through mechanisms for anonymization, abstraction, or controlled role design, represents an important direction for future research.

Supplemental Material

Supplemental Material