PLOS One study: AI debate impersonations beat politicians on authenticity and coherence

AI-generated impersonations of political figures got higher marks than the real politicians in a study published in PLOS One. Steffen Herbold of the University of Passau in Germany and colleagues report that members of the public judged AI-made debate replies as more authentic, relevant, and coherent than the speakers' actual debate responses.

That is the headline's entire shock in one sentence: the “imposter” outperformed the “original” on persuasion basics. In the study, authenticity is not about whether voters like the person, it is about whether the reply feels believable. Coherence is about whether the answer hangs together logically. Relevance is whether it addresses the moment. The finding says the public was more persuaded by the AI impersonation than by the politician’s own debate response.

Why would that happen? Even without inventing details beyond the study’s framing, the incentives around political debate help explain the opening. Debate formats reward short, confident-sounding lines and a tight argument arc. Real candidates are also constrained by strategy, time, rehearsed talking points, and the need to avoid missteps. AI systems generating impersonations, meanwhile, can be optimized toward producing fluent, structured replies that “sound like” the target figure. If the public evaluates authenticity and coherence through the lens of readability and consistency, an AI reply may simply arrive in a more polished package, even when it is not genuinely authored by the politician.

This is also a trust-and-distribution story, not just a creativity story. Political messaging now travels through layers of remixing, captioning, clips, and commentary. The modern voter often does not encounter a debate response as a full, unbroken artifact. Instead, they encounter summaries, edits, and snippets. If AI-generated impersonations are more coherent and relevant, they are more likely to survive these transformations and keep attention. The second-order effect is brutal: attention can flow to content that feels “more real” to the viewer, even when it is not.

Regulation is already trying to catch up to the capacity problem, where synthetic media becomes cheap, fast, and scalable. But regulators typically chase symptoms: disclosure requirements, provenance standards, and restrictions on deceptive political advertising. Herbold’s study points to a harder technical and governance challenge. Even if you label something as AI-generated, the viewer’s perception may still tilt toward the synthetic content if it scores higher on the judgment criteria they care about. That means compliance alone may not fix the incentive mismatch between truthfulness and perceived quality.

Boards and executives should also note the governance risk within communications teams and platforms. In organizations, approval processes exist for brand safety and reputational control. Yet the study suggests that at least some audiences evaluate “authenticity” as an output property, not an authorship property. If AI impersonations can outperform on coherence and relevance, internal teams may face a paradox: the content that looks best on persuasion metrics is exactly the content that increases verification burden and reputational exposure. That burden does not just fall on legal. It hits product, trust and safety, editorial integrity, and executive communications.

For investors and founders building in AI, this study reads like a market map with a warning label. There is likely demand for tools that can generate debate-ready responses that feel coherent and context-aware. But the same capabilities can intensify social and political instability by lowering the cost of manufacturing conviction. If “better authenticity” can be simulated, then the credibility of public discourse becomes harder to measure. The capital question becomes: can companies build effective controls, provenance, and human verification workflows that preserve trust without killing usability?

Ultimately, this is not just about politics. It is about how persuasion works in the real world. When a study finds AI impersonations judged more authentic, relevant, and coherent than actual debate responses, it suggests a future where the audience’s perceived signal quality can drift away from the genuine source. For decision-makers in media, tech, and regulated industries, the strategic stake is clear: the next battleground is not whether people can generate convincing text, it is whether institutions can maintain shared standards for what “real” means after the copy gets better than the original.