Table 1 Extracted dimensions. The table summarizes the dimensions coded from the included studies. Some dimensions were aggregated for analysis, specifically for the LLM model and domain. The interaction type “one-shot” refers to message conditions where the LLM produces a single static output, whereas “interactive” refers to multi-turn setups allowing adaptive communication.
From: A meta-analysis of the persuasive power of large language models
Dimensions | Values |
|---|---|
LLM model | GPT-3.x (i.e., GPT 3, GPT 3-davinci, GPT 3.5), GPT-4.x (i.e., GPT 4, GPT 4o mini), Claude 3.x (i.e., Claude Sonnet 3.5, Claude Sonnet 3.7) |
Interaction type | one-shot, interactive |
Domain | politics, health, other (propaganda, quiz, mixed) |
Content type | words < 200, 200 < words < 500, multiple round dialogue |
Participant type | general population, educated |
Personalization | yes, no |
Recruitment source | Prolific, mTurk, Lucid |
Outcome type | attitude, intention, behavior |
Outcome variable | attitude change, compliance rate, policy support, behavioral intention, percent agreement, perceived message effectiveness (PME) |
Outcome measurement | 7-point Likert scale, 5-point Likert scale, 101-point scale, binary proportion |
