GPT-4 Turbo Beat 112 UK Public Figures on Debate Authenticity

There is a study out this week in PLOS One that I do not think should slide past the usual 'AI can now do X' news cycle, because the specific finding is uncomfortable in a way most of these are not. A team led by Steffen Herbold at the University of Passau asked GPT-4 Turbo to impersonate 112 UK public figures, politicians, business people, journalists and others, using nothing more exotic than Wikipedia biographies and transcripts from 30 episodes of the BBC's Question Time. They then had 948 UK adults score the answers, blind, against what the real people actually said on the show.

The public preferred the AI. As 404 Media reported, the impersonated responses were rated more authentic, more coherent, and more relevant than the real debate answers. Herbold told the outlet the authenticity result was "really surprising because that's supposedly hard to fake," and the paper itself describes "a dire need to inform the general public of the potential harm this can have on society."

Why this matters for anyone doing policy, platform trust, or election work: the input the model needed was not private, was not scraped from a closed source, and did not require fine-tuning. A public bio, public debate transcripts, a general-purpose commercial model. That is a very low bar for producing text that a nationally representative sample judges as more like the person than the person's own words, during a period the researchers frame around the 2024 UK election. Herbold's recommended response is regulatory: banning political deepfakes, and educating the public on how to spot AI-generated messages.

The honest caveat is that this is one study, from one lab, on written text only, and the comparison format is a stylised debate show. It does not tell you what happens with audio, video, or long-form policy statements, and the write-up does not detail how demographic subgroups within the 948-person sample performed. The forward-looking piece is not really about defence, since the ingredients are already public. It is about whether the education and labeling response Herbold calls for actually gets built and shipped before the next election cycle needs it.