🔬 physics

From Heard to Lived Opinions: Simulating Opinion Dynamics with Grounded LLM Agents in Economic Environments

This paper introduces a novel opinion dynamics framework that grounds LLM-based agents in an economic environment, demonstrating that incorporating lived economic experiences—such as adverse conditions, inequality, and price instability—significantly shapes individual opinion rigidity and drives collective polarization and distributional shifts.

Original authors: Ryuji Hashimoto, Masahiro Kaneko, Ryosuke Takata, Takehiro Takayanagi, Kiyoshi Izumi

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ryuji Hashimoto, Masahiro Kaneko, Ryosuke Takata, Takehiro Takayanagi, Kiyoshi Izumi

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 are trying to understand why people in a society start arguing, agreeing, or becoming polarized.

Most previous studies on this topic treated people like radio stations. They assumed that if you just let these radio stations talk to each other, their opinions would change based on what they heard. If Station A says "The sky is blue," Station B might start thinking the sky is blue too.

But in real life, people aren't just radios. They are gardeners. Their opinions aren't just shaped by what they hear; they are shaped by what they do and what happens to their gardens. If a gardener tries to grow tomatoes and they all rot because of a drought, they aren't just going to listen to a neighbor say, "Don't worry, it's fine!" They are going to become angry, skeptical, and maybe even refuse to plant tomatoes again.

This paper introduces a new way to simulate human opinion using AI (Large Language Models, or LLMs). Instead of just letting the AI agents chat, the researchers put them inside a virtual economy.

Here is the breakdown of how it works, using simple analogies:

1. The Setup: The "Living" Simulation

The researchers created a digital world with 20 AI agents. Think of these agents as 20 different families living in a small town.

The Town: It has a factory (the Firm) that makes goods and a Mayor (the Government) that collects taxes and gives out basic income.
The Families: Each family has to make daily decisions: How hard should I work? What should I buy? What do I think about the town?

Unlike previous studies where the AI just "talks," here the AI has to act. If a family decides to work less, they earn less money. If they spend too much, they go into debt. The town's economy reacts to these choices: prices go up or down, wages change, and the Mayor adjusts taxes.

2. The "Grounding" Concept

The key innovation here is "Grounding."

Old Way (Un-grounded): An AI agent says, "I feel sad about the economy," just because it read a sad news article or heard a sad friend.
New Way (Grounded): An AI agent says, "I feel sad about the economy," because it actually lost its job in the simulation, its bank account dropped, and it couldn't afford groceries.

The AI's opinion is now "grounded" in its lived experience, just like a real human's opinion is grounded in their paycheck and their bills.

3. What They Discovered

The researchers ran this simulation 30 times and watched how the families' opinions changed over time. They found three fascinating things:

A. Opinions Follow the Wallet (Individual Level)

When a family did well (high income, stable prices), they were optimistic and talked about "opportunities" and "balance."
When a family struggled (low income, high prices), they became pessimistic and talked about "unfairness" and "survival."
The Metaphor: It's like a thermostat. If the room gets cold (economic hardship), the family doesn't just say they are cold; they actually act cold (turn up the heat, complain more). Their mood is directly tied to their economic reality.

B. The "Threat Rigidity" Effect

This was a surprising finding. When the economy became unstable (prices were jumping up and down wildly), the families didn't change their minds more; they changed them less.
The Metaphor: Imagine a ship in a storm. When the waves are huge, the captain doesn't try to steer left or right constantly; they lock the wheel and hold on tight.
In the simulation, when the economy was chaotic, the AI families became stubborn. They stopped changing their opinions because the threat felt too big. They clung to their existing beliefs as a safety mechanism.

C. Inequality Creates a Split (Population Level)

When the simulation created a gap between the rich families and the poor families, the town didn't just become "more unequal." It became polarized.
The Metaphor: Imagine a room where half the people are wearing gold watches and the other half are hungry. The rich people start talking about "merit and hard work," while the poor people start talking about "systemic failure." They stop understanding each other.
The simulation showed that as the gap between rich and poor grew, the two groups' opinions drifted further apart, creating two distinct camps that couldn't agree on anything.

4. Why This Matters

This paper proves that you cannot understand how societies think if you only look at what they say. You have to look at what they do and what they experience.

Before: We thought opinions were just a result of a "he said, she said" argument.
Now: We see that opinions are a result of "I earned, I spent, I survived."

By giving AI agents a "life" with real consequences (money, jobs, taxes), the researchers created a much more realistic model of human behavior. It shows that economic stress doesn't just make people poor; it makes them rigid, angry, and divided.

In short: You can't simulate human opinion in a vacuum. You have to put the humans in the game, let them play, let them win or lose, and then listen to what they have to say.

1. Problem Statement

Traditional Opinion Dynamics (OD) research and recent Large Language Model (LLM) simulations primarily focus on linguistic interactions (e.g., dialogue, framing effects, confirmation bias) to model how opinions evolve. However, these approaches often neglect lived experiences, specifically the economic consequences of individual actions. In the real world, opinions are heavily shaped by material conditions (e.g., income, employment, inflation). The absence of an economic environment in current simulations prevents the modeling of endogenous phenomena such as:

Threat-induced rigidity: Attitudes becoming more rigid under economic threat.
Inequality-driven polarization: The link between economic disparity and social polarization.

The authors argue that without grounding agents in an environment where their actions generate feedback (economic outcomes), OD simulations cannot fully capture the complexity of collective opinion formation.

2. Methodology

The authors propose a novel simulation framework where LLM-based agents are embedded within a macroeconomic environment. The system operates over $T$ time steps with $n$ agents acting as households.

A. Agent Architecture (LLM Households)

Role: Agents act as households making economic decisions (labor supply, consumption) and expressing opinions.
Decision Process: At each time step, the LLM outputs categorical actions ( $a^L, a^C \in \{low, medium, high\}$ $a^{L}, a^{C} \in {l o w, m e d i u m, hi g h}$ ) and a natural language opinion ( $O_{t,i}$ $O_{t, i}$ ).
- Categorical actions are mapped to continuous numerical values via uniform sampling within predefined intervals.
- Prompt Structure: The LLM is conditioned on:
  1. Role instructions and persona (demographics, Big Five personality traits).
  2. Public information (historical wages, prices, taxes).
  3. Observed opinions of other households.
  4. Financial History: A standardized deviation ( $z$ -score) of the agent's cash holding relative to the population mean.
  5. Previous opinion.
State Update: Cash holdings are updated based on labor income, consumption costs, taxes, and subsidies.

B. Economic Environment (Rule-Based Agents)

The environment consists of a Firm and a Government that update the state based on agent actions:

Firm: Uses a Cobb-Douglas production function ( $Y = F(K, L)$ $Y = F (K, L)$ ) to produce goods.
- Price Update: Adjusted based on the imbalance between supply ( $Y$ ) and demand ( $C$ ).
- Wage Update: Adjusted toward the Marginal Product of Labor (MPL).
Government: Implements a simplified policy:
- Basic Income: Unconditional fixed subsidy ( $b$ ).
- Progressive Taxation: Tax rates depend on the household's relative financial status (very low to very high).

C. Data Processing & Metrics

To analyze high-dimensional unstructured data, the authors discretize agent states:

Internal State ( $h_{t,i}$ ): A tuple of Sentiment (Negative, Neutral, Positive) and Financial Status (Low, Average, High).
Transition Vectors: Agents are characterized by the frequency of transitions between internal states (81 possible transition types).
Population Metrics:
- Gini Coefficient ( $G_t$ ): Measures economic inequality.
- Polarization Index ( $P^\Delta_t$ ): Measures the separation between positive and negative sentiment groups, weighted by group size.
- Opinion Instability: Measured via Wasserstein distance between opinion distributions over time.

3. Key Contributions

Grounded OD Framework: Introduced a simulation framework where LLM agents are not just conversationalists but economic actors whose opinions are shaped by the feedback loops of their own decisions.
Endogenous Opinion Dynamics: Demonstrated that opinion changes can emerge from economic interactions rather than just linguistic exchange.
Structured State Representation: Developed a method to discretize continuous economic and sentiment data into "internal states" to trace individual trajectories and cluster behavioral patterns.
Empirical Validation of Socio-Economic Coupling: Provided evidence that macroeconomic instability and inequality directly drive specific patterns in opinion dynamics (rigidity and polarization).

4. Key Results

The study addresses three Research Questions (RQs):

RQ1: Behavioral Validity:
- Agents exhibit economically rational behavior: Low-income agents increase labor supply and decrease consumption; high-income agents do the opposite.
- Opinion Grounding: Textual opinions correlate with financial states. Low-income agents with negative sentiment express themes of "unfairness" and "survival," while high-income agents with negative sentiment focus on "tax burdens" and "temporary fixes."
RQ2: Individual Opinion Dynamics:
- Heterogeneous Trajectories: Clustering revealed distinct patterns. Some agents remain stable (POS/High or NEG/Low), while others shift based on economic feedback.
- Threat-Rigidity Effect: A counter-intuitive finding: Higher price volatility leads to lower rates of sentiment change. Agents become more rigid in their opinions during economic instability, supporting the "threat-rigidity" hypothesis.
RQ3: Population-Level Dynamics:
- Inequality $\rightarrow$ Polarization: There is a statistically significant positive correlation between the Gini coefficient (inequality) and the polarization index. As economic disparity widens, the population splits into distinct opinion camps.
- Instability $\rightarrow$ Opinion Volatility: Macroeconomic volatility (price fluctuations) correlates with larger temporal shifts in the collective opinion distribution (measured by Wasserstein distance).

5. Significance and Implications

Beyond Linguistic Models: The paper proves that to accurately simulate social dynamics, AI agents must be situated in environments where they face consequences for their actions. Purely linguistic models miss critical drivers of opinion change.
Policy Insights: The simulation suggests that economic policies (e.g., managing inflation or inequality) have direct, non-linear impacts on social stability and polarization.
Future Research: This framework opens the door for "social simulacra" that can test economic policies and predict their sociological side effects before real-world implementation.

6. Limitations

Simplified Interaction: Social interactions are limited; complex processes like deliberation, persuasion networks, or voting are not modeled.
Static Institutions: The government and firm rules are fixed and do not adapt to agent opinions (no feedback from opinion to policy).
Unidirectional Influence: Currently, opinions do not directly influence economic decisions (e.g., an agent's political view does not change their labor supply), though the authors note this is a key direction for future bidirectional modeling.