The Laboratory for Comparative Social Research announces the next regular seminar, which will be held as a zoom session on September, 4, at 03:30 p.m. CET (04:30 p.m. Moscow time, GMT+3). Ulyana Shelegova, 4th year student of the "Political Science and World Politics" program will deliver a report "Exploring social and political biases in the Russian Large Language Models".

To participate, please, register via the link.

Abstract

In recent years, Large Language Models (LLMs) have transformed human-computer interaction, shifting traditional hierarchy. While the impact of LLMs on peoples’ opinion is almost indisputable, scholarly debates continue over whether they inherently favor specific ideologies.

In this study, I was aimed to explore what kinds of political and social views Yandex GPT-4 and GPT-5 tend to express and how well they manage to adapt to different cultural contexts. In order to clarify it, I  have chosen four very different cases, Sweden, Bangladesh, Serbia, and Mexico, which sit in different quadrants of the Inglehart-Welzel Cultural Map, which makes them perfect for spotting contrasts.

It turns out that both models tend to lean toward liberal-progressive answers, and this tendency grows stronger with every new generation. At the same time, GPT-5 is better at adjusting to countries where the dominant values don’t match the model’s default orientation. However it’s still not reliable enough to be taken as a true reflection of public opinion. 

Everyone interested is invited!

Working language is English.