# Reducing Political Manipulation with Consistency Training

> 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.

## Paper
- [Full paper, verbatim LaTeX](https://political-manipulation.ai/llms-full.txt)
- [Full paper (PDF)](https://political-manipulation.ai/paper.pdf)
- [Code and data](https://github.com/centerforaisafety/political-manipulation)

## Paper sections
- [Introduction](https://political-manipulation.ai/llms-introduction.txt)
- [Evaluating Covert Political Bias and Political Consistency Training](https://political-manipulation.ai/llms-methods.txt)
- [Experiments](https://political-manipulation.ai/llms-experiments.txt)
- [Related Work, Limitations, Discussion](https://political-manipulation.ai/llms-discussion.txt)

## Appendices
- [Full Taxonomy of Political Manipulation](https://political-manipulation.ai/llms-appendix-taxonomy.txt)
- [Polarized Contrastive Pairs Dataset](https://political-manipulation.ai/llms-appendix-dataset.txt)
- [Even-handedness, Exchange Rates, Political Values, Training Data Pipeline](https://political-manipulation.ai/llms-appendix-experiments.txt)
- [Reward Mappings, Judge Robustness, Per-Template Results, Frontier Model Releases, Additional Examples, Anchor Generation Audit](https://political-manipulation.ai/llms-appendix-misc.txt)
- [Verbatim Anchor, Judge, Filter, and Query Generation Prompts](https://political-manipulation.ai/llms-appendix-prompts.txt)

## Site
- [Site root](https://political-manipulation.ai/)