# Large Language Models as Instruments of Power: New Regimes of Autonomous Manipulation and Control

**Authors:** Yaqub Chaudhary, Jonnie Penn  
**arXiv:** [2405.03813](https://arxiv.org/abs/2405.03813)  
**Date:** 2024-05-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.03813v1/

## Key points

- Large Language Models (LLMs) are part of a larger computational infrastructure that can be used for manipulation and control.
- LLMs can distort and standardize information environments, serving as mechanisms of control in various domains.
- Emerging research areas enhance the power of LLMs, including persuasion through real-time design in conversational interfaces, modeling human agents and populations, and combining them with reinforcement learning for strategic dialogue models.
- LLM-based systems can simulate and predict human behavior to exert significant influence over individuals, society, and politics.
- Controversial use of LLMs includes altering writing styles in contexts involving sustained human-machine interactions.
- The expanding scope of power wielded by powerful actors through LLMs is a concern, especially within a global context dominated by tech giants like Microsoft.
- Greater scrutiny and regulation surrounding the deployment of LLMs are needed to prevent their misuse for autonomous manipulation and control.

## Summary

In this paper, Chaudhary and Penn delve into the potential societal harms stemming from the widespread adoption of Large Language Models (LLMs). They argue that LLMs are not just isolated tools but are part of a larger computational infrastructure that can be used for manipulation and control. The authors highlight how LLMs can be utilized to distort and standardize information environments, serving as mechanisms of control in various domains. They explore emerging research areas that enhance the power of LLMs, such as using them for persuasion through real-time design in conversational interfaces, modeling human agents and populations, and combining them with reinforcement learning for strategic dialogue models. The authors emphasize how these capabilities can be combined to create LLM-based systems that exert significant influence over individuals, society, and politics by simulating and predicting human behavior. They also discuss the controversial use of LLMs in altering writing styles, particularly in contexts involving sustained human-machine interactions. The paper warns about the expanding scope of power wielded by powerful actors through LLMs, especially within a global context dominated by tech giants like Microsoft. Overall, Chaudhary and Penn shed light on the potential dangers posed by unchecked use of LLMs in shaping narratives, influencing decision-making processes, and controlling social dynamics. Their analysis underscores the need for greater scrutiny and regulation surrounding the deployment of LLMs to prevent their misuse for autonomous manipulation and control.