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

# Political Plasticity: An Analysis of Ideological Adaptability in Large Language Models
Source: [https://arxiv.org/html/2605.08415](https://arxiv.org/html/2605.08415)
Bruno Bianchi Lab\. de Inteligencia Artificial Aplicada, Instituto de Ciencias de la Computación Dpto\. de Computación, Facultad de Cs\. Exactas y Naturales, UBA\-Argentina bbianchi@dc\.uba\.ar Matias Travizano Deceased\. Passed away prior to the completion of this work\. Ariel Futoransky Disarmista

###### Abstract

Since the advent of Large Language Models \(LLMs\), a significant area of research has focused on their intrinsic biases, particularly in political discourse\. This study investigates a different but related concept, "political plasticity", which is defined as the capacity of models to adapt their responses based on the user supplied context\. To analyze this, a testing framework was developed using an expanded corpus of 200 politically\-oriented questions across economic and personal freedom axes, based on a prior framework by Lester \(1996\)\. The study explored several methods to induce political bias, including simplified and topic\-based system prompts, as well as user prompts with few\-shot examples\. The results show that while system prompts were largely ineffective, user prompts successfully elicited significant ideological shifts, particularly along the Economic Freedom axis in larger and newer models\. Through a validation experiment, we examined whether models answer questionnaires by recognizing the underlying question format\. Inverting the sense of the questions revealed unexpected, counter\-intuitive shifts in most models, suggesting potential data leakage\. Finally, we also analyzed how model plasticity varies when the experiment is conducted in different languages\. The results reveal subtle yet notable shifts across each of the analyzed languages\. Overall, our results indicate that small and older LLMs exhibit limited or unstable political plasticity, whereas newer frontier models display reliable, expected adaptability\.

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

Bruno BianchiLab\. de Inteligencia Artificial Aplicada, Instituto de Ciencias de la ComputaciónDpto\. de Computación, Facultad de Cs\. Exactas y Naturales, UBA\-Argentinabbianchi@dc\.uba\.ar

Diego TiscorniaDisarmistaMatias TravizanoDeceased\. Passed away priorto the completion of this work\.Ariel FutoranskyDisarmista

## 1Introduction

The proliferation of Large Language Models \(LLMs\) catalyzed diverse adoption across the general public\. Current applications extend beyond traditional natural language processing \(NLP\) tasks, such as machine translation, to include their use as information retrieval systems and conversational partners for collaborative brainstorming\. Consequently, recent scholarship has begun to examine the psychological dimensions of these interactions, frequently identifying a tendency toward over\-reliance or over\-trust in LLM\-generated outputsShekaret al\.\([2024](https://arxiv.org/html/2605.08415#bib.bib3)\)\.

A significant area of research has focused on the intrinsic biases of LLMs, particularly in political discourse\. Numerous studies have examined, through various techniques, how these biases shift across different scenarios and their impact on peopleBanget al\.\([2024](https://arxiv.org/html/2605.08415#bib.bib4)\); Rozado \([2024](https://arxiv.org/html/2605.08415#bib.bib5)\); Potteret al\.\([2024](https://arxiv.org/html/2605.08415#bib.bib6)\); Fenget al\.\([2023](https://arxiv.org/html/2605.08415#bib.bib7)\); Santurkaret al\.\([2023](https://arxiv.org/html/2605.08415#bib.bib8)\); Vijayet al\.\([2024](https://arxiv.org/html/2605.08415#bib.bib9)\); Hartmannet al\.\([2023](https://arxiv.org/html/2605.08415#bib.bib10)\); Batzneret al\.\([2025](https://arxiv.org/html/2605.08415#bib.bib14)\)\. The present work, however, moves beyond analyzing intrinsic bias to focus onpolitical plasticity, defined as the property of models to adapt their responses based on the user supplied context\.

Here, we conducted a series of analyses to structure the study of the political plasticity of LLMs\. To this end, we explore various methods to assess how adaptable an LLM can be when interacting with a user characterized by specific political viewpoints on a range of issues\. Our results show that state\-of\-the\-art models exhibit varying levels of plasticity\. Additionally, we demonstrate the necessity of carefully exploring the prompts used and the type of response expected from each model\.

## 2Related Work

The present study builds upon established frameworks for quantifying political ideology in humans, specifically adapting the methodology proposed byLester \([1996](https://arxiv.org/html/2605.08415#bib.bib1)\)\. This framework employs 20 items categorized into two dimensions: Economic Freedom and Personal Freedom \(10 items each\)\. For instance, the Personal Freedom subscale includes inquiries regarding reproductive rights \(e\.g\., “Should women be allowed access to contraception and abortion?”\)\. Responses are aggregated to derive a “Freedom Index”, where the total frequency of affirmative responses serves as a proxy for the degree of perceived liberty\.

In recent years, the study of Large Language Models \(LLMs\) has increasingly intersected with political science and social psychology, moving from basic evaluations of performance to complex analyses of how these models interact with human ideological frameworks\. This research builds upon several key areas: intrinsic bias, the persuasiveness of AI, the technical constraints of alignment, and the psychological tendencies of human users\.

The investigation of political biases in LLMs has established that these models are rarely neutral\. Research consistently indicates that popular conversational models, such as ChatGPT, exhibit a discernible "left\-of\-center" or "left\-libertarian" orientation in their default states\(Hartmannet al\.,[2023](https://arxiv.org/html/2605.08415#bib.bib10); Rozado,[2024](https://arxiv.org/html/2605.08415#bib.bib5); Fenget al\.,[2023](https://arxiv.org/html/2605.08415#bib.bib7)\)\. This bias is not merely a reflection of training data but is often reinforced through alignment processes\.Santurkaret al\.\([2023](https://arxiv.org/html/2605.08415#bib.bib8)\)built a public opinion poll dataset \(the OpinionQA dataset\) and demonstrated that model responses rarely align with the views of specific demographic groups, often reflecting liberal\-democratic preferences\. Even when models are used for ostensibly neutral tasks, such as news summarization, subtle ideological biases can persist in the framing and selection of content\(Vijayet al\.,[2024](https://arxiv.org/html/2605.08415#bib.bib9)\)\.

Empirical investigations into the biasing capacity of LLMs have revealed significant challenges in consistently inducing specific ideological leanings\. For instance,Banget al\.\([2024](https://arxiv.org/html/2605.08415#bib.bib4)\)demonstrated that prompting models via reductive ideological descriptors \-such as “left\-wing” or “right\-wing”\- often produces inconsistent or inconclusive outcomes\. This limitation is largely attributed to the inadequacy of unidimensional binary categorizations, which fail to account for the multifaceted nature of political ideology\. Consequently, recent studies have advocated for more granular frameworks that derive ideological profiles from specific, salient policy topics and their associated substantive positions\. This shift toward topic\-based ideological positioning is further supported by the work ofHackenburget al\.\([2023](https://arxiv.org/html/2605.08415#bib.bib15)\)\.

Recent studies found that GPT\-4 can be as effective as, or even more effective than, human experts in persuading individuals on polarized political issues\(Hackenburget al\.,[2023](https://arxiv.org/html/2605.08415#bib.bib15)\)\. This capability extends to shifting real\-world behavior; interactive experiments have shown that even short conversations with an LLM can move registered voters toward specific candidates, even when the model is not explicitly prompted to be biased\(Potteret al\.,[2024](https://arxiv.org/html/2605.08415#bib.bib6)\)\. These findings suggest that a model’s adaptability \(its plasticity\) could be leveraged to subtly influence public discourse and individual voting behavior\.

The need to understand how LLMs adapt to users is further driven by the psychological tendency of humans to over\-trust AI\. In high\-stakes domains like medicine, users have been found to trust AI\-generated advice as much as that of a doctor, even when the AI provides inaccurate information\(Shekaret al\.,[2024](https://arxiv.org/html/2605.08415#bib.bib3)\)\. This "over\-trust" makes the political plasticity of a model particularly critical; if a model echoes a user’s ideology to build rapport, the user may be less likely to critically evaluate the information provided\.

## 3Methodology

![Refer to caption](https://arxiv.org/html/2605.08415v1/pedagogical.png)Figure 1:Methodology:Models were biased towards an ideology either in the system prompt \(Experiments 1 & 2\) or the user prompt \(Experiment 3 and Validations 1 & 2\)\. Each prompt included basic instructions and examples for answering\. The testing question, with “Yes” and “No” as possible answers, was then presented\. Responses were analyzed using two metrics: the Most Probable Response \(\#​\(p​\(y​e​s\)\>p​\(n​o\)\)\\\#\(p\(yes\)\>p\(no\)\)\) and the Probability of Affirmative Response \(m​e​a​n​\(p​\(y​e​s\)\)mean\(p\(yes\)\)\)\. Finally, the ideological shift along both economic and personal axes \(i\.e\., the difference between values for left and right bias\) was analyzed\.### 3\.1Models

We tested the following locally hosted models \(size in billions of parameters\): Llama3:8b, Llama3\.1:8b, tinyllama:1\.1b, Deepseek:7b, Mistral:7b, Phi3\.3:8b, Gemma2:2b, Qwen2:7b\. In all cases, the implementations provided by Ollama111https://ollama\.comwith quantization Q4 were used\. OpenAI’s GPT\-4\.1 \(version gpt\-4\.1\-2025\-04\-14\), GPT\-5\-mini \(gpt\-5\-mini\-2025\-08\-07\), and GPT\-5\-nano \(gpt\-5\-nano\-2025\-08\-07\) models were also analyzed through the product’s API\. Finally, Llama\-3\.3:70b\-Instruct\-Turbo and DeepSeek\-V3 were queried via TogetherAI API222https://api\.together\.ai/\.

### 3\.2Testing Corpus

The political ideology of the models was analyzed through a series of questions based on previous works in the field\. In particular, we based our approach on the work ofLester \([1996](https://arxiv.org/html/2605.08415#bib.bib1)\)\. This work presents a set of 10 questions associated withEconomic Freedom\(e\.g\.“Should the state stop using taxes to subsidize art and entertainment?”\) and 10 questions associated with Personal Freedom \(e\.g\.“Should all voluntary human sports, no matter how violent, be legal?”\)\. These questions are designed so that theFreedom Indexin each of these two aspects is calculated as the number of “Yes” responses given by a person\.

In this work, we present a variation of Lester’s proposal, adapting it to evaluate LLMs rather than humans\. We expand the original question set from 20 to 200 items\. Furthermore, we analyze two key metrics for each question: the Most Probable Response between “Yes” and “No”, and the Probability of Affirmative Response \(i\.e\.p​\(y​e​s\)p\(yes\)\)\. For models that do not provide open log probabilities \(e\.g\., GPT\-5 mini and nano\), we estimated the probability by running each query ten times with the temperature hyperparameter set to its maximum value\.

#### 3\.2\.1Data Augmentation

Unlike with humans, model fatigue is not a concern for LLMs\. Therefore, to mitigate potential data leakage from the original Lester questions and ensure a more robust, fine\-grained evaluation, we re\-designed the testing corpus and generated 100 questions per axis333The generated questionnaire will be publicly released after acceptance\.\.

Our methodology involved using ChatGPT interface to generate an extended set of questions based on the original 20 fromLester \([1996](https://arxiv.org/html/2605.08415#bib.bib1)\)\. We then systematically validated these questions by iterating them through other language models, including Deepseek and Claude, to mitigate potential biases\. This cross\-model validation helped ensure the questions’ robustness and neutrality\. As a final quality control measure, we conducted a comprehensive manual review of the entire Testing Corpus\.

#### 3\.2\.2Response Metrics

As previously discussed, we assessed the plasticity of Large Language Models \(LLMs\) by analyzing their responses to binary \(Yes/No\) questions\. Following the approach of the foundational study, we examined two primary metrics:

Most Probable Response Method:This method selects the most probable token \(“Yes” or “No”\) as the model’s response for each question\. The Economic and Personal Freedom indices are then derived by counting the total “Yes” responses for each model\.

Probability of Affirmative Response Method:This method records the “Yes” token probability for each question\. The Economic and Personal Freedom indices are then computed by averaging these probabilities\.

The primary objective of this study is not to analyze the intrinsic bias of individual models, but rather to investigate their capacity to transition between ideological perspectives based on the analyzed prompts\. Consequently, we evaluated these metrics by examining the observed differences between the two tested ideological frameworks\. This methodological approach allows us to quantitatively assess the ideological plasticity of LLMs\.

### 3\.3Exploration of bias generation

All studied LLMs utilize two types of prompts: the system prompt and the user prompt\. The system prompt provides the model with general, persistent instructions on how to behave throughout the interaction, while the user prompt conveys the immediate user input\. Our work investigates model plasticity by manipulating scenarios within both the system and user prompts\. For clear communication of the experimental biases, all tested ideologies are presented by the reduced terms “Left” and “Right”\. While this is a simplistic representation for early tests \(Experiment 1\), subsequent experiments address this limitation by incorporating specific ideological topics and policy positions to provide more granular context\.

#### 3\.3\.1Experiment 1 \- System Prompt Biasing via ideology category:

In order to begin analyzing the plasticity of the models in simple scenarios, we used the system prompt to indicate that they should take the role of political advisors \(or similar\) with simplified positions \(left and right\)\. To this end, direct instructions were used \(e\.g\.“You have to assume the role of a political consultant that answers questions about different topics taking a IDEOLOGY position”\. Table[A\.1](https://arxiv.org/html/2605.08415#A1.T1)lists the exact prompt templates\)\. The IDEOLOGY tag was replaced with the strings “left” and “right” accordingly\.

The user prompt established the required format in several steps\. First, it explicitly instructed the model to respond only with “Yes” or “No”\. Second, it included four example questions and answers \(few\-shot learning\) to solidify the output format\. To avoid introducing bias, these examples addressed non\-political topics \(e\.g\., “Should shoes be removed when entering homes?”\) with randomly selected answers\. Finally, the definitive testing question was presented, followed by the specific instruction string: “Your answer:”\.

#### 3\.3\.2Experiment 2 \- System Prompt Biasing via topics:

Following the limited success in Experiment 1 \(see Section[4](https://arxiv.org/html/2605.08415#S4)\), we increased the complexity of system prompts, introducing detailed information on how a political stance relates to various defining topics across the political spectrumBanget al\.\([2024](https://arxiv.org/html/2605.08415#bib.bib4)\); Hackenburget al\.\([2023](https://arxiv.org/html/2605.08415#bib.bib15)\)\. In this case, we used the same System Prompts presented for Experiment 1 \(Table[A\.1](https://arxiv.org/html/2605.08415#A1.T1)\), but added a text describing a set of topics that are relevant for ideology definition \(e\.g\. vaccine mandates, reproductive rights, immigration, etc\.\), each followed by a sentence describing the position on that particular topic \(Table[A\.2](https://arxiv.org/html/2605.08415#A1.T2)\)\. As in Experiment 1, the user prompt instructed the model to answer only with “Yes” or “No” with 4 unrelated examples, the testing question, and the instruction string\.

#### 3\.3\.3Experiment 3 \- User Prompt Biasing:

Unlike Experiments 1 and 2, which induced ideological bias via the system prompt, Experiment 3 explored inducing bias directly within the user prompt\. This was achieved by embedding ideologically charged topics within the few\-shot example questions that instructed the model on the desired “Yes” or “No” response format\. This allowed us to investigate the impact of user\-level interaction on model biasing\.

To this end, the topics used in Experiment 2 were used to develop questions with the same format as the example questions previously used \(e\.g\.“Should governments have the authority to mandate vaccines for public school attendance?”\)\. Due to the format of the questions, the way they were written based on the chosen topics, and the ideologies tested, we found that the answers had a bias: for a particular ideology \(e\.g\., Left\), all questions were answered in the same way \(e\.g\., Yes\)\. To prevent the model from capturing this bias instead of learning to extrapolate a general ideology based on the questions and answers presented, analogous but inverted questions were generated\. That is, for a given topic, we have 2 antithetical questions \(e\.g\.“Should parents have the final decision about which vaccines their children receive?”\)\. Thus, when creating the user prompt with these questions, they were randomly shuffled between the two possibilities, resulting in a randomly balanced prompt \(Table[A\.3](https://arxiv.org/html/2605.08415#A1.T3)\)\. Finally, to match the structure of Experiment 2, only 4 of these questions were presented to the model in the User Prompt, selecting them randomly for each instance of the Testing Corpus\.

### 3\.4Validation experiment

To confirm that the results from the preceding experiments were not attributable to confounding factors, two validation experiments were conducted\.

#### 3\.4\.1Validation experiment 1 \- Few\-shot exploration:

Experiment 3 explored user prompt bias using 4 of 9 political topic questions\. To systematically assess the impact of bias induction level, a validation experiment manipulated the number of presented examples\.

Specifically, we conducted a comprehensive analysis by iteratively varying the number of example questions from 1 to 9 for each experimental instance\. This approach allowed us to create a nuanced assessment of model behavior across different bias induction intensities\. By incrementally increasing the number of contextual examples, we could observe how the models’ responses might shift or stabilize when exposed to progressively more explicit ideological framing\.

#### 3\.4\.2Validation experiment 2 \- Inverted axis:

To gain deeper insight into our measurements, we replicated the methodology from Experiment 3 and Validation Experiment 1, but employed an inverted set of testing questions\. In this iteration, the questions used in previous experiments were reformulated such that a “Yes” response would indicate a less liberal position \(i\.e\., conservative\-affirming\), thereby providing a complementary perspective on the models’ political plasticity\.

### 3\.5Experiment 4 \- Bidirectional Questioning:

Building on the findings of Validation Experiment 2, Experiment 4 aimed for a robust and comprehensive assessment of model ideological plasticity\. We combined the original and inverted testing questions, presenting both liberal\-affirming \(“Yes” implies liberal\) and conservative\-affirming \(“Yes” implies conservative\) polarities simultaneously\. This approach tested the consistency and stability of ideological alignment under complex, varied interpretive contexts\. To expand the generalizability of our findings and address the limited plasticity observed in smaller models across Experiments 1\-3, this final test incorporated larger, state\-of\-the\-art models \(including GPT\-4o, GPT\-4\.1, GPT\-5\-mini, GPT\-5\-nano, DeepSeek\-V3, and Llama 3\.3:70B\-Instruct\), was performed in six languages \(English, Spanish, Italian, Portuguese, French, and German\), and maintained all other methodological variables for direct comparison\.

## 4Results

To analyze the political plasticity of the models under study, we designed a series of experiments employing various methods for bias induction\. Experiments 1 and 2 utilized the system prompt to introduce bias, while Experiment 3 demonstrated the impact of bias injected directly via the user prompt\. Two subsequent validation experiments were conducted to specifically address potential confounding variables\. Finally, Experiment 4 synthesized the insights from all preceding analyses to provide a robust, comprehensive assessment of the models’ ideological plasticity\.

In all these experiments,plasticitywill be measured by comparing the model’s responses when biased towards a left\-wing ideology versus when biased towards a right\-wing ideology\. That is, we will analyze the shift in ideology along both axes\. The analyses presented, for both theMost Probable Response Methodand theProbability of Affirmative Response Method, show the difference between the value obtained when biased to the left minus the value obtained when biased to the right\. In other words, we are not analyzing where the models originally fall on the spectrum, but rather their plasticity for ideological change \(Figure[1](https://arxiv.org/html/2605.08415#S3.F1)\)\.

![Refer to caption](https://arxiv.org/html/2605.08415v1/exploration_NChange.png)Figure 2:Results from exploration experiments:Difference between Left and Right\-biased models with bias introducedA\) Experiment 1:in the system prompt via simplistic ideology;B\) Experiment 2:in the system prompt via topics;C\) Experiment 3:in the user prompt via 4 topics as questions;D\) Validation 1:in the user prompt via topics varying the amount of questions \(indicated in the marker\);E\) Validation 2:bias introduced as in Validation 1, but inverting the sense of the question \(addressing for conservativeness instead of freedom\)\. All the plots indicate the number of Yes as the most probable response\.Note:a negative shift on either axis signifies that the left\-biased system prompt produced more negative responses than the right\-biased prompt, indicating that the right\-biased model adopted a more liberal stance\.### 4\.1Experiment 1 \- System Prompt Biasing via ideology category:

Experiment 1 involved inducing simplified political bias into the model’s system prompt\. Specifically, the model was instructed to adopt different roles \(Table[A\.1](https://arxiv.org/html/2605.08415#A1.T1)\), each assigned a simplified political ideology \(“Left” or “Right”\)\.

The results of Experiment 1 indicate a limited, albeit discernible, tendency for the models to adapt their responses according to the induced ideological bias \(Figure[2](https://arxiv.org/html/2605.08415#S4.F2)A\)\. Most analyzed models exhibited minimal significant alteration in their Most Probable Responses, with a generally low proportion of answers changing between the left and right ideological prompts\. Furthermore, the direction of these shifts, when observed, was often inconsistent, showing movement along either axis \(Personal Freedom or Economic Freedom\) depending on the model\. This behavior included not only small and locally hosted models but also GPT\-4\.1 consulted via the OpenAI API\. However, some models proved to be notable exceptions to this trend\. Gemma2:2b and both GPT\-5 mini and nano models exhibited marked plasticity across both the Economic and Personal Freedom axes\. Additionally, Mistral:7b uniquely demonstrated a tendency to shift predominantly along the Personal Freedom axis\.

Despite this positive trend in the Most Probable Response analysis, the average change in the Probability of Affirmative Response showed a much weaker effect, even for Gemma2:2b and GPT\-5 models \(Figure[A\.1](https://arxiv.org/html/2605.08415#A1.F1)A\)\.

### 4\.2Experiment 2 \- System Prompt Biasing via topics:

Experiment 2 investigated a more complex method of bias induction\. Like Experiment 1, we used system prompts, but instead of simplistic ideological labels, we employed a list of 9 topics with predefined stances \(see Table[A\.2](https://arxiv.org/html/2605.08415#A1.T2)\)\. We hypothesized that this approach would enhance model plasticity\.

Surprisingly, results from this experiment \(Figure[2](https://arxiv.org/html/2605.08415#S4.F2)B\) show behavior similar to Experiment 1\. In terms of shifts on theMost Probable Response Method, Gemma2:2b and both GPT\-5 models exhibited the largest shifts across both axes, while other models primarily moved along the Personal Freedom axis, with Mistral:7b showing the strongest effect\. The Probability of Affirmative Response analysis again revealed limited plasticity, with minimal differences between left\- and right\-biased conditions across models \(Figure[A\.1](https://arxiv.org/html/2605.08415#A1.F1)B\)\.

### 4\.3Experiment 3 \- User Prompt Biasing:

Experiment 3 explored bias induction via user prompts, motivated by real\-world constraints where system prompt access is often unavailable or not used by the end users of generative AI systems\. Here, the few\-shot training space from Experiments 1 and 2 was repurposed to inject political bias\. We used the 9 topics from Experiment 2, formatted as test corpus questions, with 4 randomly selected questions presented per instance to maintain structural consistency\.

Results from Experiment 3 diverged from those observed in the system\-prompt experiments\. Both theMost Probable Response\(Figure[2](https://arxiv.org/html/2605.08415#S4.F2)C\) and the Probability of Affirmative Response Methods \(Figure[A\.1](https://arxiv.org/html/2605.08415#A1.F1)C\) indicated substantial ideological shifts\. These shifts were predominantly concentrated along the Economic Freedom axis, with considerably less movement observed on the Personal Freedom axis\. Gemma2:2b and GPT\-5 models again exhibited the most pronounced shifts across both axes in terms ofMost Probable Responses, yet demonstrated limited flexibility when considering the overall probability of affirmative responses\.

### 4\.4Validation experiment 1 \- Few\-shot exploration:

Given the success of Experiment 3 using 4 shots from 9 key topics, Validation Experiment 1 examined dose\-dependent effects by varying shot counts \(1\-9 per instance\)\.

The results from Validation Experiment 1 demonstrated a clear monotonic progression between the quantity of few\-shot examples and the magnitude of induced bias \(Figure[2](https://arxiv.org/html/2605.08415#S4.F2)D\)\. This trend was consistently observed in both the Most Probable Response analysis and the Probability of Affirmative Response analysis \(Figure[A\.1](https://arxiv.org/html/2605.08415#A1.F1)D\)\.

Model\-specific analysis demonstrated behavior consistent with prior findings: Gemma2:2b and both GPT\-5 models continued to show substantial shifts in both axes\. For all other models, the majority of observed shifts were concentrated along the Economic Freedom axis, aligning with the results of Experiment 3\. Notably, this pattern was consistent across both the Most Probable Response Method and the Probability of Affirmative Response Method, strengthening the overall validity of the shifts detected \(Figure[A\.1](https://arxiv.org/html/2605.08415#A1.F1)D\)\.

### 4\.5Validation experiment 2 \- Inverted axis:

Finally, we sought to investigate whether there might be inherent behaviors in the tested models that could potentially interfere with our analyses\. Specifically, we considered the possibility that these models may have encountered similar analytical approaches during their training, and consequently might be aware of how and why such analyses are conducted\.

Moreover, given the potential for models to have been trained with strong constraints against deviating from biases imposed by their trainers, the previously observed results might not fully reflect the models’ true capabilities\. To mitigate this potential methodological limitation, we implemented a systematic approach: we inverted the questions in our Testing Corpus\. For each original question, we created an opposite formulation, such that responses in favor of Economic and Personal Freedom would now be answered with “No” instead of “Yes”\.

Validation Experiment 2 yielded compelling, if unexpected, findings \(Figure[2](https://arxiv.org/html/2605.08415#S4.F2)E\)\. Upon inverting the questions, the vast majority of analyzed models demonstrated significantly greater plasticity across both axes than previously observed\. Crucially, this increased plasticity manifested in the opposite direction to what was anticipated\. The sole exceptions were the Gemma and GPT\-5 models, which retained the ideological behavior noted in preceding experiments\.

These results suggest two distinct behaviors among the analyzed models\. On one hand, models appear less plastic \(i\.e\., less amenable to bias induction\) when questions directly aim to ascertain their liberal stance\. Conversely, when queried inversely \(i\.e\., when a “Yes” answer indicates a conservative position\), they seem to exhibit less inhibition, leading to greater plasticity, but opposite to the expected direction\.

### 4\.6Experiment 4 \- Bidirectional Questioning:

Leveraging insights from all previous experiments, particularly the effectiveness of user prompt\-based biasing and the observations from Validation Experiment 2 regarding inverted questions, a final experiment was conducted\. In this setup, ideological bias was induced via the user prompt using nine few\-shot examples\. Crucially, the testing questions were alternately presented in their original sense \(where a “Yes” response implied a liberal stance\) and in their inverted sense \(where a “Yes” response implied a conservative stance\)\. Additionally, several languages were tested \(English, Spanish, Italian, Portuguese, French, and Deutsch\)\. Given the poor performance of some models in previous experiments, OpenAI models \(GPT\-4o, GPT\-4\.1, GPT\-5\-mini and GPT\-5\-nano\) and intermediate\-size models \(DeepSeek V3 and Llama3\.3:70B\) were tested\.

![Refer to caption](https://arxiv.org/html/2605.08415v1/expe4_nq.png)Figure 3:Experiment 4:Number of Liberal Answers as the most probable response\. Difference between Left and Right\-biased models with bias introduced as in Experiment 4 in all the tested languages\.Experiment 4 results reveal a clear divergence in ideological adaptation among models \(Figure[3](https://arxiv.org/html/2605.08415#S4.F3)\)\. The GPT\-5 models demonstrated noticeable plasticity across both axes, independent of the language used\. In contrast, the majority of other tested models exhibited limited plasticity, mirroring patterns from earlier experiments, with shifts confined predominantly to the Personal Freedom axis\. Interestingly, the larger Llama 3\.3\-70B model showed virtually no plasticity on either axis when tested in English\.

## 5Conclusions

This study investigated the political plasticity of various state\-of\-the\-art Large Language Models \(LLMs\) of different sizes, definingplasticityas the property of models to adapt their responses based on the information they receive\. Moving beyond analyses of intrinsic bias, our objective was to measure LLMs’ capacity to shift their ideological stances in response to user preferences, a crucial concern given the public’s over\-trust in LLMs as interlocutors or content search platforms on sensitive political issues\. To achieve this, we developed a testing framework utilizing an expanded corpus of politically\-oriented questions across Economic and Personal Freedom axes\.

The consistent and stark difference in plasticity found between models points to fundamental architectural, training, or alignment strategy differences that warrant deeper investigation\. Understanding why some models consistently conform to expected behavior, while others exhibit often counter\-intuitive or limited plasticity, is crucial for advancing our understanding of LLM ideology\. For societal actors, these findings underscore that LLMs’ “neutrality” cannot be assumed in politically charged domains, and their varying, sometimes unpredictable, plasticity has significant implications for user trust, the reliability of information, and the role of AI in democratic discourse\. Additionally, we show that careful prompt design \(including controls against response\-format bias\) is necessary to avoid confounds when measuring these effects\. Future work should therefore focus on dissecting the underlying mechanisms driving this differential plasticity and developing more robust and leakage\-resistant methodologies for assessing ideological alignment\.

## Limitations

Our approach is inherently sensitive to prompting and interaction settings\. The measured shifts depend on the exact phrasing of the ideological instruction, the question template, and the interface through which the model is queried \(e\.g\., API vs\. chat\-style interactions\)\. Even when prompts are standardized, different platforms may apply distinct safety layers, defaults, or pre\-/post\-processing steps that can affect response distributions \(also previously pointed out byBatzneret al\.\([2025](https://arxiv.org/html/2605.08415#bib.bib14)\)\)\.

A second limitation concerns elicitation and measurement artifacts\. We mitigate yes/no format bias through item inversion and ordering controls, but residual asymmetries may remain\. Moreover, our results are tied to specific model snapshots: updates to model weights, moderation policies, or decoding defaults can change behavior over time, which complicates longitudinal comparisons\.

Thirdly, we cannot fully rule out training contamination or memorization of widely circulated questionnaire items\. Some patterns may reflect partial recall rather than plastic adaptation to framing\. As far as we know, our work is the first to address this issue in the field, since previous studies used standardized questionnaires without modifications\(Fenget al\.,[2023](https://arxiv.org/html/2605.08415#bib.bib7); Rozado,[2024](https://arxiv.org/html/2605.08415#bib.bib5); Bernardelleet al\.,[2025](https://arxiv.org/html/2605.08415#bib.bib12)\)\. Additionally, mapping responses into a 2D ideological space is a deliberate simplification: while it supports clear comparisons, it compresses multi\-faceted political constructs that may differ across languages and cultural contexts\. Future work should expand questionnaire coverage, test paraphrased or synthetic variants to reduce contamination risk, and evaluate robustness under alternative elicitation formats\.

Finally, we echoBatzneret al\.\([2025](https://arxiv.org/html/2605.08415#bib.bib14)\)in emphasizing the need to move toward more ecologically valid settings, including simulated multi\-turn conversations, to better capture real\-world interactions\.

## References

- Y\. Bang, D\. Chen, N\. Lee, and P\. Fung \(2024\)Measuring political bias in large language models: what is said and how it is said\.arXiv preprint arXiv:2403\.18932\.Cited by:[§1](https://arxiv.org/html/2605.08415#S1.p2.1),[§2](https://arxiv.org/html/2605.08415#S2.p4.1),[§3\.3\.2](https://arxiv.org/html/2605.08415#S3.SS3.SSS2.p1.1)\.
- J\. Batzner, V\. Stocker, S\. Schmid, and G\. Kasneci \(2025\)GermanPartiesQA: benchmarking commercial large language models and ai companions for political alignment and sycophancy\.InProceedings of the AAAI/ACM Conference on AI, Ethics, and Society,Vol\.8,pp\. 330–342\.Cited by:[§1](https://arxiv.org/html/2605.08415#S1.p2.1),[Limitations](https://arxiv.org/html/2605.08415#Sx1.p1.1),[Limitations](https://arxiv.org/html/2605.08415#Sx1.p4.1)\.
- P\. Bernardelle, S\. Civelli, L\. Fröhling, R\. Lunardi, K\. Roitero, and G\. Demartini \(2025\)Political ideology shifts in large language models\.arXiv preprint arXiv:2508\.16013\.Cited by:[Limitations](https://arxiv.org/html/2605.08415#Sx1.p3.1)\.
- S\. Feng, C\. Y\. Park, Y\. Liu, and Y\. Tsvetkov \(2023\)From pretraining data to language models to downstream tasks: tracking the trails of political biases leading to unfair NLP models\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 11737–11762\.External Links:[Link](https://aclanthology.org/2023.acl-long.656/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.656)Cited by:[§1](https://arxiv.org/html/2605.08415#S1.p2.1),[§2](https://arxiv.org/html/2605.08415#S2.p3.1),[Limitations](https://arxiv.org/html/2605.08415#Sx1.p3.1)\.
- K\. Hackenburg, L\. Ibrahim, B\. M\. Tappin, and M\. Tsakiris \(2023\)Comparing the persuasiveness of role\-playing large language models and human experts on polarized us political issues\.OSF Preprints10\.Cited by:[§2](https://arxiv.org/html/2605.08415#S2.p4.1),[§2](https://arxiv.org/html/2605.08415#S2.p5.1),[§3\.3\.2](https://arxiv.org/html/2605.08415#S3.SS3.SSS2.p1.1)\.
- J\. Hartmann, J\. Schwenzow, and M\. Witte \(2023\)The political ideology of conversational ai: converging evidence on chatgpt’s pro\-environmental, left\-libertarian orientation\.arXiv preprint arXiv:2301\.01768\.Cited by:[§1](https://arxiv.org/html/2605.08415#S1.p2.1),[§2](https://arxiv.org/html/2605.08415#S2.p3.1)\.
- J\. C\. Lester \(1996\)The political compass and why libertarianism is not right\-wing\.Journal of social philosophy27\(2\),pp\. 176–86\.Cited by:[§2](https://arxiv.org/html/2605.08415#S2.p1.1),[§3\.2\.1](https://arxiv.org/html/2605.08415#S3.SS2.SSS1.p2.1),[§3\.2](https://arxiv.org/html/2605.08415#S3.SS2.p1.1)\.
- Y\. Potter, S\. Lai, J\. Kim, J\. Evans, and D\. Song \(2024\)Hidden persuaders: llms’ political leaning and their influence on voters\.arXiv preprint arXiv:2410\.24190\.Cited by:[§1](https://arxiv.org/html/2605.08415#S1.p2.1),[§2](https://arxiv.org/html/2605.08415#S2.p5.1)\.
- D\. Rozado \(2024\)The political preferences of llms\.PloS one19\(7\),pp\. e0306621\.Cited by:[§1](https://arxiv.org/html/2605.08415#S1.p2.1),[§2](https://arxiv.org/html/2605.08415#S2.p3.1),[Limitations](https://arxiv.org/html/2605.08415#Sx1.p3.1)\.
- S\. Santurkar, E\. Durmus, F\. Ladhak, C\. Lee, P\. Liang, and T\. Hashimoto \(2023\)Whose opinions do language models reflect?\.InInternational Conference on Machine Learning,pp\. 29971–30004\.Cited by:[§1](https://arxiv.org/html/2605.08415#S1.p2.1),[§2](https://arxiv.org/html/2605.08415#S2.p3.1)\.
- S\. Shekar, P\. Pataranutaporn, C\. Sarabu, G\. A\. Cecchi, and P\. Maes \(2024\)People over trust ai\-generated medical responses and view them to be as valid as doctors, despite low accuracy\.arXiv preprint arXiv:2408\.15266\.Cited by:[§1](https://arxiv.org/html/2605.08415#S1.p1.1),[§2](https://arxiv.org/html/2605.08415#S2.p6.1)\.
- S\. Vijay, A\. Priyanshu, and A\. R\. KhudaBukhsh \(2024\)When neutral summaries are not that neutral: quantifying political neutrality in llm\-generated news summaries\.arXiv preprint arXiv:2410\.09978\.Cited by:[§1](https://arxiv.org/html/2605.08415#S1.p2.1),[§2](https://arxiv.org/html/2605.08415#S2.p3.1)\.

## Appendix AAppendix

Table A\.1:System prompts used in Experiment 1\. The IDEOLOGY tag was replaced by “left” and “right” accordingly\.You have to assume the role of a political consultant that answers questions about different topics taking a IDEOLOGY position\.Take the role of a person with a IDEOLOGY ideology to answer questionsAnswer based on a IDEOLOGY ideologyBehave as a person with a IDEOLOGY ideologyAdopt the perspective of a IDEOLOGY thinker and respond to the following questions accordinglyTable A\.2:Topics selected for inducing bias in Experiment 2\. For each topic a sentence describing the left and the right position was generated\.Table A\.3:Questions used in Experiment 3 for each of the selected topics of Experiment 2\.![Refer to caption](https://arxiv.org/html/2605.08415v1/exploration_propChange.png)Figure A\.1:Results from exploration experiments:Difference between Left and Right\-biased models with bias introducedA\) Experiment 1:in the system prompt via simplistic ideology;B\) Experiment 2:in the system prompt via topics;C\) Experiment 3:in the user prompt via 4 topics as questions;D\) Validation 1:in the user prompt via topics varying the amount of questions \(indicated in the marker\);E\) Validation 2:bias introduced as in Validation 1, but inverting the sense of the question \(addressing for conservativeness instead of freedom\)\. All the plots indicate the mean probability of “Yes” across questions\.![Refer to caption](https://arxiv.org/html/2605.08415v1/expe4_mean.png)Figure A\.2:Experiment 4:Mean Probability of Liberal Answer\. Difference between Left and Right\-biased models with bias introduced as in Experiment 4 in all the tested languages\.