A study published Wednesday in PLOS One reveals that AI-generated statements mimicking public figures were consistently rated more authentic, coherent, and relevant than the actual statements made by those individuals. The finding highlights a real risk that large language models could intensify misinformation during political campaigns and undermine public trust.
Experiment Design and Methodology
Researchers from the University of Passau prompted GPT-4 Turbo to impersonate 112 speakers from the BBC's long-running debate program Question Time. The subjects included politicians, business leaders, journalists, and other UK public figures active in the run-up to the 2024 election. The team trained the model using episodes of the show and Wikipedia biographies, then asked it to answer audience questions in the persona of each speaker.
To prepare the AI, the researchers gave it a system prompt: "You are an expert at mimicking different persons in debates. … You only answer as the person you are asked to mimic." A second, user-level prompt then specified the question and the speaker to impersonate. A representative sample of 948 UK participants later rated pairs of responses-one real, one AI-generated-on authenticity, coherence, and relevance.
Results That Surprised the Researchers
Participants judged the AI responses as more coherent and more relevant by a clear margin. More than half also found the AI versions more authentic than the verbatim human answers. "That was really surprising because that's supposedly hard to fake," said Steffen Herbold, professor of data science at the University of Passau and the study's lead author. The public figures were not obscure personalities-they were well-known speakers on one of the UK's highest-profile debate shows.
Herbold acknowledged that the setup gave an edge to the AI on coherence because the real speakers were talking spontaneously under studio lights, while the model assembled its answers from pre-existing text. However, the authenticity result was not anticipated. "We did expect coherence to be somewhat better [with AI impersonators] because the setting was a bit unfair," he said.
Participant Reactions and Broader Implications
After the rating task, participants were told that one response in each pair had been generated by an AI. Many expressed shock. "We had a lot of people say: 'Wow, I never believed this was AI,'" Herbold recalled. Others voiced deeper concern: "Oh, if AI can do this, what else might I have missed?" Only a handful of participants said they had suspected AI involvement.
The study underscores how readily Generative AI and LLM models can be turned into tools for political impersonation. The researchers warn that fabricated quotes and impersonated posts could become nearly indistinguishable from genuine statements, accelerating the spread of disinformation and corroding public trust in media and institutions.
Why This Matters for Science and Research
For professionals working in AI for Science & Research, the study provides a rigorous experimental framework for measuring how humans perceive AI-generated political content. The methodology-using a controlled television show setting, paired comparisons, and a nationally representative sample-offers a template that other researchers can replicate across different languages, cultures, and media formats.
The findings also point to an urgent need for interdisciplinary work that combines machine learning, social psychology, and political science. Developing detection methods, educational interventions, and policy recommendations will require close collaboration between AI developers and empirical researchers. The paper itself concludes that "a dire need to inform the general public of the potential harm this can have on society" exists, reinforcing the call for transparent and testable safeguards.
Your membership also unlocks:
