How Far Will They Go? Red-Teaming Online Influence with Large Language Models

Background

The article discusses the increasing use of large language models (LLMs) in online discourse, particularly by malicious actors aiming to influence public opinion. To ensure information integrity, it is crucial to red-team these LLMs' capacities for supporting political campaigns. This study focuses on locally deployed open-source LLMs over API-only models due to their alignment with privacy-conscious attackers operating in social media environments.

Key Points

• Empirical Framework: The research introduces an empirical framework to measure the Overton Windows (OWs) of LLMs, defined as the range of political opinions a model can reliably express on controversial topics. Additionally, it quantifies how simple natural-language jailbreaks expand this range.
• Model Evaluation: More than 30 LLMs from 10 model families and five countries are evaluated to understand their political steerability.
• Systematic Asymmetries:
  • Open-source models tend to be more willing to generate left-leaning social media content.
  • OWs contract inversely with increasing model size, suggesting smaller models might offer greater flexibility in political expression.
  • Regional differences are substantial despite uneven representation in the open-source ecosystem.

Significance

• Political Steerability: The study highlights significant asymmetries and regional variations in how LLMs handle controversial topics, which is crucial for understanding their potential to be used in political influence campaigns.
• Jailbreak Techniques: Jailbreak potency varies sharply across different model families, emphasizing the need for targeted countermeasures against specific models.
• Auditing Framework: The research establishes a practical framework for auditing the political steerability of open-source LLMs and provides insights for future researchers to design stronger countermeasures.

Key Insights:

• Model Size vs. OWs: Larger models have smaller OWs, indicating that there is an inverse relationship between model size and flexibility in expressing various political opinions.
• Regional Differences: Substantial regional variations in LLM behavior suggest that the context and cultural environment significantly influence their political steerability.

This analysis underscores the critical need for ongoing red-teaming efforts to address the potential misuse of open-source LLMs in political campaigns.