💬 NLP

When Do LLMs Generate Realistic Social Networks? A Multi-Dimensional Study of Culture, Language, Scale, and Method

This study demonstrates that Large Language Models generate realistic social networks whose structural properties and demographic biases are significantly shaped by prompt design, cultural framing, language, and model scale, revealing that these technical choices function as substantive sociological variables rather than mere implementation details.

Original authors: Sai Hemanth Kilaru, Sriram Theerdh Manikyala, Raghav Upadhyay, Sri Sai Kumar Ramavath, Srivika Nunavathu, Dalal Alharthi

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sai Hemanth Kilaru, Sriram Theerdh Manikyala, Raghav Upadhyay, Sri Sai Kumar Ramavath, Srivika Nunavathu, Dalal Alharthi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a digital "matchmaker" (a Large Language Model, or LLM) and you ask it to introduce 50 fictional people to each other to see who becomes friends. You might think, "If I just ask the computer nicely, it will create a realistic map of how humans actually connect."

This paper puts that idea to the test. The researchers treated the AI not just as a tool, but as a social scientist. They asked: Does the way we ask the AI to make friends change the kind of friend group it creates?

Here is what they found, explained through everyday analogies:

1. The "How You Ask" Matters (The Prompt Architecture)

The researchers tried four different ways to ask the AI to build these networks. Think of these as different social scenarios:

Sequential: Asking each person, "Who do you want to be friends with?" (Like a roll call).
Global: Showing the AI the whole list of 50 people at once and saying, "Connect them all." (Like a party host looking at the whole room).
Local: Telling the AI, "Only look at these 5 people nearby and decide." (Like a small group conversation).
Iterative: Letting the AI make connections, then asking it to update them based on the new group. (Like a game of musical chairs where friendships evolve).

The Finding: These aren't just technical tweaks; they are like changing the rules of a game.

When the AI looked at people one by one (Sequential), political views became the biggest reason people became friends.
When the AI looked at the whole list at once (Global), it got distracted by surface details. Instead of politics, age became the main reason for friendship. It's like if you were in a huge crowd, you might notice someone's gray hair before you notice their political bumper sticker.

2. The "Where You Ask" Matters (Culture)

The researchers asked the AI to simulate friend groups in four different cultural settings: the US, India, Japan, and Brazil. Even though the list of 50 people (the "roster") stayed exactly the same, the culture of the prompt changed the result.

The Finding: The "glue" holding the group together changed. In some cultures, the group was very tightly knit (everyone knew everyone); in others, it was more fragmented. The AI didn't just copy-paste the same social structure; it adapted the "vibe" of the network based on the cultural context you gave it.

3. The "Size of the Brain" Matters (Model Scale)

They tested three versions of the AI: a tiny one, a medium one, and a full-sized one.

The Finding: You might think the tiny AI just makes "noisier" or "sloppier" versions of the big AI's work. It doesn't.
The tiny AI produces a fundamentally different kind of network. It's not just a blurry photo of the big AI's picture; it's a completely different drawing. If you swap a big AI for a small one to save money, you aren't just getting a cheaper version of the same result; you are getting a different social reality entirely.

4. The "Language You Speak" Matters (Prompt Language)

They kept the culture and the people the same but changed the language of the prompt (English, Spanish, Hindi, Japanese).

The Finding: This was the most surprising part.

Politics: The AI's view on political friendship was stubborn. No matter if you asked in Hindi or Japanese, the AI still thought politics was the main driver of friendship.
Religion: This was the opposite. When the AI was asked in Hindi, it suddenly became obsessed with religion as a reason for friendship. When asked in Japanese, it focused more on age.
The Analogy: It's like the AI has a "spotlight." In English, the spotlight shines on politics. In Hindi, the spotlight swings over to religion. The language you speak changes what the AI thinks is important, even if the people in the room haven't changed.

5. Are These Fake Networks "Realistic"?

Finally, they compared the AI's fake networks to real human networks and to old-school math formulas used to simulate networks (like random chance or "rich get richer" models).

The Finding:

The Good News: The AI's networks look more like real human groups than the old math formulas do. They have the right amount of "clustering" (friends of friends being friends) and "modularity" (distinct groups forming).
The Bad News: The AI is too good at creating "echo chambers." It creates groups where people are friends with others who are exactly like them (same politics, same religion) much more often than real humans do.
The Trap: Because the networks look so structurally realistic (they pass the "math test"), people might trust them. But they are hiding a secret: they are amplifying biases. The AI is building a world where everyone is more segregated and polarized than they actually are in real life.

The Bottom Line

The paper concludes that LLMs are not neutral mirrors. They are like a camera with a very specific, invisible filter.

If you change the lens (how you ask), you change who gets connected.
If you change the location (culture), you change how tight the group is.
If you change the language, you change which traits (religion vs. politics) matter most.

The "neutrality" of the AI is an illusion. The choices researchers make about how to prompt the AI are actually making deep sociological decisions about what kind of society they are simulating.

Technical Summary: When Do LLMs Generate Realistic Social Networks?

Problem Statement
Large Language Models (LLMs) are increasingly utilized as substitutes for human subjects in behavioral simulations, including the generation of synthetic social networks. While prior research indicates that LLMs produce structurally realistic graphs, the specific dependencies of their relational outputs on prompt design, cultural framing, prompt language, and model scale remain poorly characterized. Existing studies have not sufficiently isolated how these variables interact to shape network topology and homophily (the tendency to associate with similar others). Furthermore, the validity of using LLMs as relational proxies is contested, with concerns regarding systematic biases that may not be apparent in individual-level outputs but emerge in aggregate network structures.

Methodology
The study formalizes LLM-based network generation as a family of distinct conditional distributions over edge sets, indexed by four specific prompting architectures derived from sociological theories of tie formation:

Sequential: The model acts as an ego responding to a roster (factorizing by source node).
Global: The model acts as an omniscient observer assigning ties to the full roster simultaneously (non-factorizing, joint sampling).
Local: The model operates within a bounded context of a curated neighborhood (factorizing by source, restricted context).
Iterative: The model updates ties via a Markov chain over edge sets (non-factorizing, dynamic).

The experimental design involves a fixed roster of 50 demographically grounded personas (gender, age, race/ethnicity, religion, political affiliation, interests) sampled from U.S. Census marginals. The study generates 192 verified directed networks across three independent experiments:

Cultural Study: Varies cultural framing (US, India, Japan, Brazil) with fixed English language and sequential method.
Method Study: Varies the four prompting architectures across the four cultural contexts.
Language Study: Fixes culture to the US and varies prompt language (English, Spanish, Hindi, Japanese) across all four methods.

All experiments utilize three GPT-4.1 model variants (nano, mini, full) at a fixed sampling temperature ( $\tau = 0.8$ ) with two seeds per condition. Network realism is evaluated against classical baselines (Erdős/Rényi, Barabási/Albert, Watts/Strogatz) using metrics including density, clustering coefficient, modularity, and the proportion of nodes in the largest connected component (LCC). Homophily is measured using the baseline/inbreeding decomposition, distinguishing between ties induced by group size and residual preference.

Key Results

Cultural Framing: Cultural context produces measurable, non-uniform shifts in network structure. Political affiliation is the most culture-sensitive dimension under sequential, local, and iterative methods. The Largest Connected Component (LCC) is the topology metric most sensitive to cultural framing, indicating that culture affects global cohesiveness more than edge counts.
Dominant Tie-Formation Drivers: Political affiliation dominates tie formation in three of the four methods (sequential, local, iterative) across all cultures and languages, consistently exceeding random baselines. However, the Global method acts as a systematic exception, where age becomes the dominant homophily dimension, likely due to attentional degradation over long contexts where surface features are more salient than subtle political signals.
Model Scale Divergence: Models are not interchangeable. A stable divergence ranking exists across all studies: $d(\text{full}, \text{mini}) < d(\text{mini}, \text{nano}) < d(\text{full}, \text{nano})$ . The smallest variant (nano) produces a qualitatively different network structure rather than merely a noisier version of larger models.
Prompt Language Asymmetry: Varying prompt language (with culture fixed) shifts religion homophily significantly (especially under Hindi prompting, where the ratio peaks at 1.74), while political homophily remains remarkably invariant across all four languages. This suggests that language acts as a soft selector for specific demographic axes based on training data co-occurrence patterns.
Structural Realism vs. Bias: LLM-generated networks match real social graphs better than classical baselines on clustering (0.61 vs. 0.45 for real networks) and modularity (0.50 vs. 0.38). However, they encode demographic biases (specifically inbreeding homophily) that exceed empirically observed levels.

Significance and Claims
The paper argues that the apparent neutrality of LLM-based social simulation is illusory. Prompt design choices, often treated as implementation details, encode substantive sociological assumptions.

Prompt Architecture as a Sociological Variable: The choice of generation method (e.g., sequential vs. global) fundamentally alters which demographic dimensions drive tie formation.
Language Sensitivity: Prompt language is not a neutral conduit; it selectively amplifies specific homophily dimensions (like religion under Hindi) while leaving others (like politics) invariant.
Model Scale Implications: Substituting smaller models for larger ones for cost reasons does not yield a "noisier" version of the same network but a structurally distinct one, challenging the assumption that scaling alone resolves demographic distortions.
Methodological Recommendation: The authors contend that structural diagnostics (clustering, modularity, homophily decomposition) and sensitivity analyses across prompt architectures and languages must be standard reporting requirements for LLM-based simulations. Treating these variables as fixed defaults risks baking in unexamined demographic priors, particularly as LLM-generated networks are increasingly used as training or evaluation data for downstream models, potentially creating feedback loops of synthetic bias.