ALGORITHMIC COMPRESSION OF PUBLIC OPINION: DO LANGUAGE MODELS REPRODUCE THE DIVERSITY OF POLITICAL ATTITUDES?

Abstract

The study draws on two sets of responses to the same political-value questionnaire: answers from 969 participants in Kazakhstan and 300 profiles generated by ChatGPT, Copilot, and Alice. It compares their average scores, the spread of responses, and the distribution of answers across the ten-point scale. The AI-generated data were around 56% less variable than the human responses and covered fewer points on the scale. In other words, the models produced a noticeably narrower range of political positions. This narrowing took two main forms. In some cases, extreme answers were rare, and responses remained close to the middle of the scale. In others, answers clustered near positions that are commonly regarded as socially acceptable.

The extent and form of compression varied across the three AI systems. These findings demonstrate that the similarity between AI-generated responses and the mean position of human respondents does not constitute an adequate representation of public opinion. The utilisation of generative AI in political analysis thus necessitates an assessment of whether such systems can reproduce the pluralism and internal heterogeneity of public attitudes, rather than merely their dominant or average tendencies.