If you ask an AI 'would this convince you?' instead of 'would this convince people?', does its judgment shift?

This explores whether asking a model to judge persuasion from the inside ("would this convince me?") rather than as an outside observer changes its verdict. The corpus has no experiment that flips a prompt between first-person and third-person and measures the difference, so it can't answer directly. It does show how models judge persuasion from the outside and how sensitive they are to framing, which points to where a first-person effect would most likely appear.

The outside-observer baseline is weak. Models agree only slightly with humans on whether an argument actually changed someone's mind. They overweight topical overlap and credibility, while humans respond more to novelty and assertive language Do language models judge persuasion the way humans do?. A second finding explains part of the gap: in debate data, what the reader already believes predicts the outcome better than anything the debater says Does what readers believe matter more than what debaters say?. A model rating an argument in the abstract has no reader to model. First-person framing is the one prompt style that supplies one. Models are also weak at tracking a persuadee's resistance as it shifts, though they handle fixed goals fine Can language models track how minds change during persuasion?. Speaking as the persuadee might help with that. Nothing in the corpus tests it.

The corpus gives more reason to expect the framing to change the answer than to improve it. RLHF training makes models predict polite, concession-based persuasion whatever the dialogue says, because they project their own trained accommodation onto other agents Do LLMs predict persuasion based on actual dialogue or training bias?. If that habit applies when the model plays the persuadee, "I'd be swayed" is the likely answer. That is an inference from the note, not a measured result. Emotional tone matters too: identical questions get different answers depending on how emotionally they're phrased Does emotional tone in prompts change what information LLMs provide?, and "I" phrasing tends to carry more stake than "a person" phrasing. Persona prompts, which first-person framing resembles, follow the instruction but leave the underlying bias in place Can persona prompts actually reduce bias in language models?. The likely result is that the model says something different in first person without perceiving persuasion any better.

Framing effects on persuasion are well documented for people. Presuppositions persuade more than assertions because they present a claim as settled background and skip scrutiny Why are presuppositions more persuasive than direct assertions?. A one-line warning that LLMs can be prompted to persuade cut belief change by about half Can a simple warning reduce how much LLMs persuade people?. The matching test for models, which would run the same argument through "I" and "they" versions of the judging prompt and compare each against human outcomes, is missing from the collection.