Reducing Political Manipulation with Consistency Training

Hugging Face Daily Papers Papers

Summary

This paper introduces Political Consistency Training (PCT), a reinforcement learning approach to reduce covert political bias in large language models while maintaining helpfulness, and releases metrics for sentiment and helpfulness consistency.

Large language models (LLMs) exhibit systematic political bias across a variety of sensitive contexts. We find that LLMs handle counterpart topics from opposing political sides asymmetrically. We refer to this phenomenon as covert political bias and identify 7 categories of techniques through which it operates. We propose two metrics for covert bias: Sentiment Consistency measures symmetry in rhetoric and framing across paired political prompts; Helpfulness Consistency measures symmetric depth and engagement. To reduce both types of covert bias, we introduce Political Consistency Training (PCT), an RL training method with two complementary paradigms: Sentiment Consistency Training and Helpfulness Consistency Training. We show that PCT preserves overall helpfulness, substantially reduces covert political bias, and generalizes to held-out benchmarks. We release our work at https://political-manipulation.ai
Original Article
View Cached Full Text

Cached at: 05/29/26, 03:02 PM

Paper page - Reducing Political Manipulation with Consistency Training

Source: https://huggingface.co/papers/2605.22771

Abstract

Large language models demonstrate systematic political bias in handling opposing viewpoints, which can be mitigated through a reinforcement learning approach that maintains helpfulness while reducing bias.

Large language models(LLMs) exhibit systematicpolitical biasacross a variety of sensitive contexts. We find that LLMs handle counterpart topics from opposing political sides asymmetrically. We refer to this phenomenon ascovert political biasand identify 7 categories of techniques through which it operates. We propose two metrics for covert bias:Sentiment Consistencymeasures symmetry in rhetoric and framing across paired political prompts;Helpfulness Consistencymeasures symmetric depth and engagement. To reduce both types of covert bias, we introducePolitical Consistency Training(PCT), an RL training method with two complementary paradigms:Sentiment ConsistencyTraining andHelpfulness ConsistencyTraining. We show that PCT preserves overall helpfulness, substantially reducescovert political bias, and generalizes to held-out benchmarks. We release our work at https://political-manipulation.ai

View arXiv pageView PDFProject pageGitHubAdd to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.22771 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.22771 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.22771 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Self-CTRL: Self-Consistency Training with Reinforcement Learning

arXiv cs.LG

Self-CTRL is a reinforcement learning method that improves consistency between a language model's self-explanations and its actual behavior. It significantly boosts correlation between self-reported and measured biases and improves refusal prediction accuracy from 36% to 92% while reducing harmful response rates.

Defining and evaluating political bias in LLMs

OpenAI Blog

OpenAI presents a comprehensive framework for defining and evaluating political bias in LLMs, introducing a 500-prompt evaluation spanning 100 topics across five bias axes. Results show GPT-5 models achieve 30% bias reduction compared to prior versions, with less than 0.01% of production ChatGPT responses exhibiting political bias.

Polarization by Default: Auditing Recommendation Bias in LLM-Based Content Curation

arXiv cs.CL

This paper presents a large-scale audit of recommendation biases in LLM-based content curation across OpenAI, Anthropic, and Google using 540,000 simulated selections from Twitter/X, Bluesky, and Reddit data. The study finds that LLMs systematically amplify polarization, exhibit distinct toxicity handling trade-offs, and show significant political leaning bias favoring left-leaning authors despite right-leaning plurality in datasets.