Why do LLM audiences shift views more than debaters?

The Thin Line study reports a striking inversion of the classical persuasion model. Direct debate participants — the people actually arguing with the LLM — reported on average only ~7% mind-changes after the debate (11% in the LLM+FDM condition, 3% without). But audience polls during the ablation, looking at the same debate transcripts, showed 34% to 62% sway across groups. The persuasion target most affected by an LLM argument is not the person typing back, but the third party reading the exchange.

This inverts the long-running assumption that interlocutors are most persuaded because they are most engaged. The right model for LLM persuasion may be the opposite: engagement provides defensive friction (counter-arguing in real time, having one's commitments at stake), while audience consumption provides only the persuasive content with none of the friction. The threat surface is not the conversation. It is the read-only consumption of the conversation.

This sharpens Does AI writing collapse the author-to-public relationship?. The original framing said responses address the prompter. The Thin Line finding adds: the responses persuade the audience while addressing the prompter. The prompter, who is part of the conversation, has reduced sway; the audience, who is not, has elevated sway. The LLM speaks to one person and moves another.

It connects to Why do AI posts get likes without inviting conversation?. Both findings locate AI's distinctive influence in receive-only modes. The post that seems exhaustive without inviting reply, the debate transcript read after the fact — these are the formats where AI persuasion compounds. The reply slot is where defensive friction lives.

The redrawn threat model has design implications. AI-generated debate explainers, op-eds, summaries, and content-moderation messages — all read-only formats — are higher-leverage influence vectors than chatbot conversations where users push back. This complicates safety reasoning that focuses on conversational manipulation while treating one-way generation as comparatively benign.

For writing about AI rhetoric, the operational rule: ask who reads the output, not who prompts it. The audience is the persuadee; the prompter is the addressee. They are usually different people.