Study: GPT-4 Turbo impersonations of 112 politicians beat the real thing

AI-generated impersonations of politicians didn't just pass as real. They outperformed the real answers on the exact quality people use to judge credibility. Researchers at the University of Passau used GPT-4 Turbo to impersonate 112 public figures in the run-up to the 2024 UK election, presented the outputs next to genuine debate answers, and found participants rated the synthetic responses higher on authenticity, coherence and relevance.

What the study actually tested

The mechanism matters more than the headline. This was not a deepfake video or a voice clone. It was text: an LLM prompted to answer as a named public figure, judged blind against that figure's real words.

• 112 public figures impersonated by the model.
• Timing tied to the UK 2024 election cycle, a live political context.
• Participants scored AI answers higher than real ones on authenticity, coherence and relevance.

The last point is the one to sit with. "Authenticity" is supposed to be the thing a fake fails. Here it's where the model won.

Why the AI scored higher

There's a plain explanation, and it isn't magic. Real debate answers are hedged, interrupted, off-topic and shaped by the pressure of the moment. An LLM produces the clean, on-message version a listener expects a politician to say. People aren't detecting truth. They're pattern-matching to a mental template of "what a competent politician sounds like," and the model fits that template better than the human does.

That reframes the risk. The danger is not that AI mimics a specific person's style perfectly. It's that fluent, coherent, plausible text reads as more trustworthy regardless of whether it's genuine.

Why it matters for RO/EE

Eastern Europe is a high-value target for exactly this class of cheap text manipulation. The region runs consequential elections on tight timelines, has active foreign influence operations, and hosts many smaller-language media markets where fact-checking capacity is thin. Romanian, Hungarian, Bulgarian and other CEE languages are well within GPT-4-class capability now.

The cost structure is the story for anyone building or investing here:

• Text impersonation is near-free to produce at scale. No video budget, no voice samples, no editing.
• Detection is the harder, fundable problem. Provenance, watermarking and source-attribution tooling are where the defensible product surface sits — not another content generator.
• Regulation is coming to meet this. The EU AI Act and election-integrity rules will push platforms and campaigns toward provenance requirements, creating a compliance market for CEE startups working in local languages that global vendors under-serve.

The recent trend line in this coverage has been AI capability compounding — trading agents, model share shifts, new labs shipping. This is the same curve pointed at information integrity. The models got good enough to be persuasive before the tooling to verify them got deployed.

Key takeaways

• GPT-4 Turbo impersonations of 112 figures beat real answers on authenticity — meaning fluency, not accuracy, drives perceived credibility.
• The exploit is cheap text, not deepfakes — the defensive opportunity is provenance and detection, especially in smaller CEE languages.
• Election-timed context makes this operational, not theoretical — CEE's dense electoral calendar and thin local fact-checking raise the exposure.