💰 quantitative finance

Talking Politics with Artificial Intelligence

Based on an analysis of 4.3 million human-AI conversations, this paper finds that AI primarily serves as a practical intermediary for routine political tasks rather than a public forum for expression, though major events like the 2024 U.S. election can temporarily shift user behavior toward more expressive and ideologically charged interactions.

Original authors: Ziwen Zu

Published 2026-07-02✓ Author reviewed ⓘ

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ziwen Zu

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 by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a new, incredibly smart digital assistant living in your pocket. For years, we've wondered: Is this assistant a town square where people shout their opinions, or is it more like a quiet library where people go to get things done?

This paper, "Talking Politics with Artificial Intelligence," takes a massive look at 4.3 million real conversations between humans and AI (like ChatGPT) to answer that question. Here is what the researchers found, explained simply:

1. The "Library" vs. The "Town Square"

The study found that for most people, talking to AI about politics is not like posting on social media. It's more like going to a librarian for help.

The Town Square (Social Media): People go there to perform, argue, show off their identity, and rally the crowd.
The Library (AI): People go there to ask, "What does this law mean?" "Can you draft a letter to my city council?" or "Summarize this 50-page government form for me."

The Big Finding: About 90% of political conversations with AI are practical. Users are asking for information, drafting documents, or translating confusing government jargon. Only a tiny sliver (less than 10%) of these chats are people actually arguing or stating their personal political opinions.

2. The "Secret" Political Life

The paper suggests that AI is acting as a political intermediary. Think of it as a translator between the average citizen and the confusing world of government.

Just as a real-life translator helps you navigate a foreign country, AI helps people navigate taxes, benefits, court rulings, and immigration forms.
This happens mostly in private. It's a "one-on-one" conversation, not a broadcast. The user isn't trying to convince an audience; they are trying to solve a problem.

3. The "Election Day" Surprise

The researchers wanted to see if a big political event would change how people use AI. They looked at the 2024 U.S. Presidential Election.

Before the winner was announced: People used AI mostly for practical things, just like before.
The Moment the Winner was Called: The moment news broke that Donald Trump had won, something shifted specifically for American users.
- Suddenly, American users started using AI to express opinions more often.
- Their language became more emotional (angry, excited, or "us vs. them").
- Their views became more extreme.

Crucially: This change only happened for people in the U.S. People in other countries who were watching the news on the same day did not change how they used AI. They kept asking practical questions. This proves that the shift wasn't just because "news was happening"; it was because the result had a direct, emotional impact on the Americans involved.

4. Where Does This Happen?

The study also looked at where these conversations happen:

Private Chats (WildChat/LMSYS): These are mostly practical. People are doing homework for their civic duties.
Shared Chats (ShareChat): When people choose to make their AI chat public (like sharing a link), the conversation changes. It becomes more like a town square, with more opinions and arguments.
Global Differences: In countries with less access to reliable news or where government forms are very confusing, people use AI for politics more often than in wealthy countries. In these places, AI acts as a lifeline to get basic information that is hard to find elsewhere.

The Bottom Line

The paper concludes that AI is not replacing the public square. Instead, it has become a civic utility.

Most of the time: It's a tool for the quiet, boring, but necessary work of citizenship (filling out forms, understanding rules).
During big moments: When a major event hits (like an election result), it briefly transforms into a place for emotional expression and partisan identity, but only for the people directly affected by that event.

In short: We mostly use AI to do politics (get things done), and only occasionally to feel politics (argue and express).

Technical Summary: Talking Politics with Artificial Intelligence

Problem and Theoretical Framework
This paper investigates the role of Large Language Models (LLMs) in political communication, specifically addressing whether these systems function as new arenas for public political expression or as conversational intermediaries for routine political demand. Existing literature largely treats AI models as objects of evaluation (e.g., sources of bias or persuasion) or analyzes their outputs. This study shifts focus to the "demand side": the user prompts themselves. The authors posit two competing theoretical possibilities:

Expressive Politics: If LLMs function like social media, usage should resemble public-facing dynamics (opinion, argument, identity signaling).
Civic Intermediary: If LLMs function as tools for navigating the state, usage should be practical (explanation, summarization, drafting, document processing).

The paper argues that LLMs occupy a unique space in political communication: private, task-oriented, and dyadic, distinct from the audience-facing nature of social media. The central research question is how political content manifests in this private, conversational environment and how it responds to major political events.

Methodology
The study utilizes a dataset of 4.30 million human–AI conversations from three public corpora: WildChat (consented public ChatGPT interactions with geolocation), LMSYS-Chat (model-comparison platform interactions), and ShareChat (user-shared conversation links). The data spans 2023 to 2025.

Measurement Strategy: The unit of analysis is the user message. The authors employ a structured, LLM-assisted coding procedure using GPT-4.1-mini and Claude-Haiku-4.5 to classify messages.
- Political Content: Defined broadly to include elections, policy, law, public administration, and institutional navigation, excluding fictional or "office politics."
- Use Cases: Categorized into information seeking, writing/drafting, document processing, opinion expression, mobilization, and institutional help.
- Ideology: Measured only for messages where the user explicitly expresses a position, scored on economic and social left-right axes (−2 to +2) and affective intensity (0 to 3).
Validation: The automated classifiers were validated against a human-coded benchmark of 3,000 messages. The political/non-political classifier achieved 84.9% accuracy, 0.77 precision, and 0.92 recall. Reliability between the two LLM coders was high (Cohen's $\kappa = 0.80$ for political classification).
Causal Design: To test the responsiveness of AI conversation to political events, the authors employ a Regression-Discontinuity-in-Time (RDiT) design centered on the Associated Press (AP) call of the 2024 U.S. presidential election (November 6, 2024). This design compares U.S. conversations immediately before and after the result call, using conversations outside the U.S. as a geographic control group to isolate the effect of the specific partisan outcome from general global news attention.

Key Results

Prevalence and Nature of Political Use:
- Political content appears in 3.9% of conversations overall (ranging from 3.3% in WildChat to 18.7% in ShareChat).
- Political use is predominantly practical. Across political user turns, 64.7% seek information/explanation, 13.0% involve writing/drafting, and 11.8% involve document processing. Only 9.4% involve opinion or argument.
- Publicness matters: ShareChat (shared links) shows significantly higher rates of opinion expression (25.0%) compared to non-shared corpora, suggesting that the presence of an audience shifts usage from assistance to expression.
- Depth matters: Political prevalence increases monotonically with conversation length, rising from ~2.6% in single-turn exchanges to over 12% in longer threads, indicating that political engagement often requires iterative problem-solving.
Geographic Variation:
- Political share is higher in lower-income and politically constrained regions (e.g., Sub-Saharan Africa at 5.8%) compared to wealthy democracies like North America (2.3%). This suggests AI may serve as a substitute for scarce or censored political information in constrained environments.
Ideological Expression:
- Explicit stance-taking is rare overall (8.2% of political conversations in WildChat).
- When stance is taken, economic positions lean mildly left, while social positions lean mildly right. Affective intensity is moderate, and ideological extremity is present but not dominant.
Event Response (2024 U.S. Election):
- Attention vs. Expression: The AP result call increased the prevalence of political conversation globally (including outside the U.S.), consistent with agenda-setting theory.
- Expressive Shift: However, the character of the conversation changed significantly only for U.S. users. Post-call, U.S. political conversations saw a 8.5 percentage point increase in stance-taking and a 6.4 point increase in affective charge.
- Geographic Specificity: No comparable shift occurred in non-U.S. conversations, ruling out platform-wide artifacts or general news shocks.
- Mechanism: The increase in expression was not solely due to a spike in election-topic prompts; the effect persisted across non-election political topics (policy, immigration, etc.), suggesting the result acted as a cue that activated expressive politics across domains.

Significance and Contributions
The paper makes two primary contributions:

Methodological: It develops and validates a procedure for measuring political content, use cases, and ideological expression in large, multilingual human–AI corpora. It demonstrates that these traces reveal a form of political communication (private, task-oriented demand) largely missed by surveys and social media data.
Theoretical: It repositions LLMs as civic intermediaries rather than just public squares. The findings suggest that citizens primarily use AI to navigate the administrative and informational burdens of the state (interpreting policies, drafting appeals, translating rules). However, the system is capable of shifting toward expressive, partisan, and affective dynamics when major political events make stakes explicit for the user.

The authors conclude that LLMs absorb routine civic demand but register the partisan and affective shocks of public events, operating as a dual-purpose interface that is practical by default but expressive under specific political conditions.

1. The "Library" vs. The "Town Square"

2. The "Secret" Political Life

3. The "Election Day" Surprise

4. Where Does This Happen?

The Bottom Line

More like this