Researchers from the University of Passau have studied how convincingly generative AI can mimic political statements. The result: responses formulated by a language model in the style of real politicians were often perceived by test subjects as more authentic, relevant, and coherent than the original statements by the politicians themselves.

For the study, the research team led by Prof. Steffen Herbold and Prof. Annette Hautli-Janisz analyzed more than 500 responses from the BBC debate show Question Time. New responses in the style of the respective politicians were generated using ChatGPT-4 Turbo. In a representative survey with 948 participants from Great Britain, the subjects did not know whether they were evaluating a real or AI-generated statement.

The results show that AI-generated responses were perceived on average as more authentic, coherent, and thematically relevant. At the same time, more than half of the AI responses differed in content from the original statements - yet they were often considered credible. The researchers see this as a significant risk for political communication and the spread of disinformation.

The study highlights how powerful modern language models have become in imitating linguistic style and patterns of argumentation. From the research team's perspective, this underscores the need to clearly label AI-generated political content and to better inform the public about the possibilities and risks of generative AI. More than 85 percent of the respondents called for the use of AI in political contributions to be disclosed.