Authoritarian Reality Simulation Using Large Language Models

In June 2024, Tencent, a Chinese multinational tech giant, released a paper that describes a new method to creating one billion synthetic personas based on internet data using Large Language Models. According to the authors, the created personas can be used to simulate reality, in which governments can use the synthetic personas to test “new policies, radical initiatives, and social dynamics” and gain insights before implementation. In this sense, LLMs not only help tackle the information insufficiency faced by many authoritarian regimes, but also create a way to test and improve their control and manipulation measures with minimal cost. However, it is uncertain how these LLM-generated personas learn the attitudes and preferences within the training data, which may consist of data generated from democratic contexts like the US or from authoritarian contexts like China. These predispositions can have a material impact on the behaviour of these personas. Against this background, we will empirically evaluate how well do LLMs simulate authoritarian realities. In particular, we will address the following research questions: RQ1. How do personas vary in their observed behaviour as a function of specified nationality? RQ2. Do personas from authoritarian contexts behave differently from democratic contexts? RQ3. How well do these variations among personas associate with known variations in socio-political attitude across democratic and authoritarian countries? To address the research questions, we will create a set of personas for each of the 90 countries covered by the World Values Survey using the framework developed by Tencent. To address RQ1 and RQ2 we will create an opinion survey about various policies and measure the variations of simulated behaviour from the personas across nationalities. To address RQ3, the responses from the personas will be compared against the results of the World Values Survey to identify possible associations.