# Incivility and Rigidity: Evaluating the Risks of Fine-Tuning LLMs for Political Argumentation

arXiv: https://arxiv.org/abs/2411.16813 · HTML: https://mlsanity.com/paper/2411.16813/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 5 of its 20 references that are in the map.
- First posted on arXiv 25 Nov 2024.

## Builds on (most cited first, 5 of 5)

- [A Note on Shumailov et al. (2024): `AI Models Collapse When Trained on Recursively Generated Data'](https://mlsanity.com/paper/2410.12954/) (2024)
- [Moral Mimicry: Large Language Models Produce Moral Rationalizations Tailored to Political Identity](https://mlsanity.com/paper/2209.12106/) (2022)
- [Argue with Me Tersely: Towards Sentence-Level Counter-Argument Generation](https://mlsanity.com/paper/2312.13608/) (2023)
- [AQuA -- Combining Experts' and Non-Experts' Views To Assess Deliberation Quality in Online Discussions Using LLMs](https://mlsanity.com/paper/2404.02761/) (2024)
- [Critical Questions Generation: Motivation and Challenges](https://mlsanity.com/paper/2410.14335/) (2024)

## Built on by (most cited first, 1 of 1)

- [How Far Did They Go? The Persuasive Tactics of Covert LLM Agents in a Discontinued Field Experiment](https://mlsanity.com/paper/2606.05256/) (2026)

## Abstract (arXiv)

Incivility on platforms such as Twitter (now X) and Reddit complicates the development of AI systems that can support productive, rhetorically sound political argumentation. We present experiments with textitGPT-3.5 Turbo fine-tuned on two contrasting datasets of political discourse: high-incivility Twitter replies to U.S. Congress and low-incivility posts from Reddit's textitr/ChangeMyView. Our evaluation examines how data composition and prompting strategies affect the rhetorical framing and deliberative quality of model-generated arguments. Results show that Reddit-finetuned models generate safer but rhetorically rigid arguments, while cross-platform fine-tuning amplifies adversarial tone and toxicity. Prompt-based steering reduces overt toxicity (e.g., personal attacks) but cannot fully offset the influence of noisy training data. We introduce a rhetorical evaluation rubric - covering justification, reciprocity, alignment, and authority - and provide implementation guidelines for authoring, moderation, and deliberation-support systems.