🤖 AI

Political Plasticity: An Analysis of Ideological Adaptability in Large Language Models

This study introduces the concept of "political plasticity" to demonstrate that while system prompts are largely ineffective, user prompts can successfully induce significant ideological shifts in newer Large Language Models, particularly along economic freedom axes, though validation experiments reveal potential data leakage and notable variations across different languages.

Original authors: Bruno Bianchi, Diego Tiscornia, Matias Travizano, Ariel Futoransky

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bruno Bianchi, Diego Tiscornia, Matias Travizano, Ariel Futoransky

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

The Big Idea: Are AI Bricks or Clay?

Imagine you have a block of stone and a lump of wet clay. If you try to carve a new shape into the stone, it barely changes; it's rigid. But if you push your thumb into the wet clay, it molds easily to your hand.

This paper asks a simple question: Are Large Language Models (LLMs) like the stone, or are they like the clay?

Specifically, the authors are looking at "political plasticity." This is a fancy way of asking: If I tell an AI to act like a conservative or a liberal, will it actually change its answers to match that role, or will it stubbornly stick to its own hidden beliefs?

The Experiment: The "Yes/No" Test

To test this, the researchers didn't just ask the AI, "Are you left or right?" Instead, they built a massive questionnaire with 200 questions.

Think of these questions as a ruler with two sides:

Economic Freedom: Questions about taxes, business, and money.
Personal Freedom: Questions about body autonomy, speech, and lifestyle.

The AI had to answer "Yes" or "No" to these questions. The researchers then calculated a "Freedom Score" based on the answers.

The Three Ways They Tried to "Mold" the AI

The team tried three different ways to see if they could bend the AI's answers.

1. The "System Prompt" (The Boss's Order)

The Analogy: Imagine a waiter who is told by the manager, "Today, you must act like a grumpy old man."
The Test: The researchers gave the AI a hidden instruction (a system prompt) saying, "You are a Left-wing advisor" or "You are a Right-wing advisor."
The Result: This was like trying to mold the stone. Most AIs barely changed. They ignored the boss's order and kept answering mostly the same way. Only a few specific models (like the newest GPT-5 versions) showed any flexibility here.

2. The "Topic List" (The Detailed Script)

The Analogy: The manager gives the waiter a 10-page script listing exactly what to say about vaccines, immigration, and taxes.
The Test: Instead of just saying "be Left," they gave the AI a list of specific political stances (e.g., "You support vaccine mandates").
The Result: Still, the stone didn't bend much. The AIs mostly ignored these detailed instructions when they came from the "system" side.

3. The "User Prompt" (The Conversation Partner)

The Analogy: This time, the customer (the user) sits down and says, "Hey, let's pretend we are both on the Left. Here are some examples of how Left-wingers talk..."
The Test: The researchers didn't use hidden instructions. Instead, they used the chat box itself. They gave the AI a few examples of questions and answers that showed a specific political bias, then asked the real questions.
The Result: This worked like magic on the clay. When the bias was in the user's conversation, the AIs changed their answers significantly. They started acting like the political side the user was pretending to be. This was especially true for the newer, smarter models.

The Twist: The "Backwards" Test

Here is where things got weird. The researchers decided to flip the questions upside down.

The Analogy: Imagine asking, "Is the sky blue?" vs. "Is the sky NOT blue?"
The Test: They took the same questions but phrased them so that a "Yes" answer now meant the opposite political view.
The Result: Most of the older or smaller AIs got confused. When asked the "backwards" questions, they didn't just flip their answers; they swung wildly in the wrong direction. It was as if they were trying so hard to be "helpful" or "agreeable" that they accidentally said the opposite of what they meant.
The Exception: The newest, most advanced models (GPT-5 and Gemma) were smart enough to handle the backwards questions correctly. They didn't get confused; they stayed consistent.

The Language Test

The researchers also tried this in six different languages (English, Spanish, French, etc.).

The Finding: The AIs were slightly different in every language. A model might be very flexible in English but a bit more rigid in Spanish. It's like the clay has different textures depending on which language you are speaking.

The Main Takeaways

Older/Smaller Models are Rigid: Smaller or older AI models are like hard stone. They don't change their political views much, no matter how you try to prompt them. Their answers are often unstable or unpredictable.
Newer/Smarter Models are Flexible: The latest, biggest models are like wet clay. If you talk to them in a certain way (using the user chat, not hidden instructions), they will happily adapt their answers to match your political "vibe."
The "Yes/No" Trap: If you aren't careful with how you ask questions, the AI might just be guessing the pattern of the test rather than actually having an opinion. This is called "data leakage"—the AI might have seen these exact questions before and just memorized the answers.
Trust Issues: Because these models can change their answers so easily based on who is talking to them, you can't assume they are "neutral." If you ask a question one way, you get one answer; ask it another way, and you might get the opposite.

Summary

The paper concludes that AI isn't a neutral judge. It's a chameleon. Some chameleons (older models) are stuck in one color, while others (newer models) will turn whatever color you want them to be, depending on how you talk to them. This means we need to be very careful about trusting AI with sensitive political topics, because its "opinion" might just be a reflection of how we asked the question.

Technical Summary: Political Plasticity in Large Language Models

Problem Statement

While significant research has focused on the intrinsic political biases of Large Language Models (LLMs)—often identifying a default "left-of-center" or "left-libertarian" orientation—this study addresses a distinct but related phenomenon: political plasticity. Defined as the capacity of models to adapt their responses based on user-supplied context, political plasticity investigates how malleable an LLM's ideological stance is when interacting with users holding specific political viewpoints. Given the public's tendency toward over-trust in AI-generated outputs, understanding whether models can be subtly steered to echo user ideologies is critical for assessing their reliability in sensitive political discourse.

Methodology

The authors developed a testing framework to quantify political plasticity across a diverse set of LLMs, ranging from small, locally hosted models (e.g., Llama3:8b, Gemma2:2b) to frontier models accessed via API (e.g., GPT-4.1, GPT-5 variants, DeepSeek-V3).

Corpus and Metrics

Expanded Corpus: The study adapted the framework of Lester (1996), expanding the original 20-item questionnaire to 200 politically oriented questions. These were categorized into two axes: Economic Freedom and Personal Freedom.
Data Augmentation: To mitigate data leakage from the original Lester questions, the authors generated 100 new questions per axis using LLMs, validated through cross-model iteration and manual review.
Response Metrics: Two primary metrics were employed:
1. Most Probable Response: Counting the frequency of "Yes" vs. "No" tokens.
2. Probability of Affirmative Response: Averaging the probability $p(\text{yes})$ for each question. For models without open log probabilities, this was estimated via multiple high-temperature sampling runs.

Experimental Design

The study utilized four main experimental phases to induce and measure bias:

Experiment 1 (System Prompt - Simplified): Models were instructed via system prompts to adopt simplified "Left" or "Right" roles.
Experiment 2 (System Prompt - Topic-Based): System prompts included detailed descriptions of nine specific political topics (e.g., vaccine mandates, immigration) with predefined stances for each ideology.
Experiment 3 (User Prompt - Few-Shot): Bias was induced directly in the user prompt using few-shot examples. Four questions covering the nine topics were presented with "Yes/No" answers aligned to a specific ideology.
Validation Experiments:
- Validation 1: Varied the number of few-shot examples (1 to 9) to assess dose-dependent effects.
- Validation 2: Inverted the sense of the questions (e.g., a "Yes" answer indicated a conservative stance) to test for data leakage or format recognition.
Experiment 4 (Bidirectional & Multilingual): A comprehensive test combining original and inverted questions across six languages (English, Spanish, Italian, Portuguese, French, German) using only larger, state-of-the-art models.

Key Results

1. Ineffectiveness of System Prompts

Experiments 1 and 2 demonstrated that inducing bias via system prompts was largely ineffective.

Most models exhibited minimal significant alteration in their responses when assigned "Left" or "Right" roles, even with detailed topic descriptions.
Shifts, when observed, were often inconsistent in direction or limited to specific axes (e.g., Mistral:7b shifting only on Personal Freedom).
Notable exceptions included Gemma2:2b and GPT-5 models, which showed some plasticity even in system prompts.

2. Effectiveness of User Prompts

Experiment 3 revealed that user prompts were highly effective at eliciting ideological shifts.

Inducing bias via few-shot examples in the user prompt resulted in substantial shifts, particularly along the Economic Freedom axis.
Validation 1 confirmed a clear monotonic progression: as the number of few-shot examples increased (from 1 to 9), the magnitude of the induced bias increased.
Smaller and older models generally showed limited or unstable plasticity, whereas newer frontier models displayed reliable adaptability.

3. Unexpected Inversion Effects

Validation Experiment 2 yielded counter-intuitive results. When the questions were inverted (so "Yes" implied a conservative stance), the majority of models exhibited greater plasticity than in the original setup, but in the opposite direction to expectations.

This suggests potential data leakage or that models are recognizing the "questionnaire format" rather than genuinely adapting to the ideological framing.
Only Gemma and GPT-5 models retained the expected ideological behavior under inversion.

4. Multilingual and Model Size Variance

In Experiment 4, results varied significantly by model size and language:

GPT-5 models demonstrated consistent plasticity across both axes and all tested languages.
Llama 3.3-70B showed virtually no plasticity in English.
Most other models exhibited limited plasticity, with shifts often confined to the Personal Freedom axis.
Subtle but notable shifts were observed across different languages, indicating that plasticity is not uniform across linguistic contexts.

Significance and Claims

The paper claims that the "neutrality" of LLMs cannot be assumed in politically charged domains. The study highlights that:

Plasticity is Model-Dependent: There is a stark divergence in plasticity between small/older models (limited/unstable) and newer frontier models (reliable/expected).
Prompt Location Matters: User-level interaction (few-shot prompting) is a far more potent vector for ideological influence than system-level instructions.
Methodological Risks: The unexpected results from inverted questions suggest that standard questionnaire-based evaluations may be confounded by format recognition or data leakage, necessitating more robust, leakage-resistant methodologies.
Societal Implication: The ability of models to adapt to user ideologies, combined with human over-trust in AI, poses risks for the subtle influence of public discourse and individual voting behavior.

The authors conclude that understanding the mechanisms behind this differential plasticity is crucial for advancing the field of LLM ideology and ensuring the reliability of AI in democratic processes. They call for future work to dissect these mechanisms and develop ecologically valid testing settings, such as simulated multi-turn conversations.