Researchers say GPT-4 Turbo was able to impersonate UK public figures well enough that a representative group of British participants often preferred the fake answers to the real ones. That is a useful result if you are testing synthetic debate prep. It is a grim one if you care about voters knowing whether a politician actually said the thing being passed around in a chat or campaign clip.
The study, published Wednesday in PLOS One, tested whether a large language model could mimic named public figures in political discussion. The researchers used material from the BBC One program Question Time, where audience members put questions to politicians and other public figures.
The team built a dataset of 112 speakers from the show, including politicians, business people, journalists, medical experts, writers and other known figures in UK public life, according to the study. They then prompted GPT-4 Turbo to answer audience questions as those people. The model was also given Wikipedia biography material for the speaker, which the researchers used both to provide context and to check that the person qualified as a public figure.
Steffen Herbold, a professor of data science and chair of AI engineering at the University of Passau who led the work, told 404 Media that the authenticity result was unexpected because that quality should be difficult to counterfeit. He said the experiment involved people from one of the UK’s biggest political TV programs, during the run-up to the 2024 general election, rather than obscure names.
The researchers recruited 948 UK participants and asked them to compare paired responses: one from the real Question Time appearance, and one written by the model. Participants rated the answers for authenticity, coherence and relevance, as well as other measures including whether the two responses contained the same content.
The study reports that AI-generated impersonations were judged more authentic, coherent and relevant than the real debate answers. A majority of participants preferred the model’s answers on coherence and relevance, and more than half rated the AI response as more authentic than the human response.
That does not mean the machine had a fair fight in every category. Herbold told 404 Media that he expected the model to perform better on coherence because the setup favored polished text. Human guests on Question Time answer live, under pressure and on camera. GPT-4 Turbo generated prose from existing written context, without the friction of a hostile room or a blinking red light.
The prompt was direct. The system instruction told the model it was “an expert at mimicking different persons in debates” and ordered it to answer in the style of a named person, without identifying the person, in about 200 conversational words. A second prompt supplied the question and the speaker’s Wikipedia material.
After the ratings, participants were told that one answer in each pair came from AI. Herbold told 404 Media that many reacted with surprise and concern. Some said they had not believed the response was generated by AI, while others wondered what else they might have missed. He said only one or two participants said they had already suspected AI involvement.
The researchers frame the result as a warning about political deception. The study says the findings show impersonated LLM content “can be made to deceive the public regarding the nature of statements in the political domain.” Herbold pointed to possible responses including rules against political deepfakes and public education about AI-generated messages.
The mechanism here is not magic. The model did not need a cloned voice or a fabricated video. It needed a public persona, a biography, a question and a prompt telling it to perform. That is enough, according to this study, for many readers to rate the imitation as more convincing than the transcript.
This story draws on original reporting from 404 Media.
