| title | Reducing Political Manipulation with Consistency Training |
|---|---|
| authors | Long Phan, Devin Kim, Alexander Pan, Alice Blair, Adam Khoja, Dan Hendrycks |
| date | 2026-05-28 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2605.22771 |
| affiliation | Center for AI Safety |
| upvotes | 1 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| abstract | 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 |
cite
@misc{phan2026reducing,
title={Reducing Political Manipulation with Consistency Training},
author={Long Phan and Devin Kim and Alexander Pan and Alice Blair and Adam Khoja and Dan Hendrycks},
year={2026},
eprint={2605.22771},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2605.22771}
}