Abstract
The ethical and responsible use of conversational artificial intelligence in public communication has become increasingly important because interactive systems can strengthen understanding and persuasion while raising concerns about transparency, user autonomy, and the responsible governance of artificial-intelligence-mediated communication. Previous research has not sufficiently established whether conversational artificial intelligence differs from static online media in counter-attitudinal climate communication in ways that are relevant to ethical standards, information credibility, and responsible persuasive influence. The study aims to examine whether conversational artificial intelligence, relative to static online media, is associated with differences in perceived understanding, argument persuasiveness, influence on perspective, and explicit self-reported opinion change, while accounting for initial attitude certainty. The analysis uses a non-randomized quasi-experimental online survey of 100 respondents conducted in 2026, with 50 participants exposed to ChatGPT and 50 to static online media, across four environmental topics: veganism, nuclear energy, recycling, and electric vehicles. The statistical analysis combines Welch independent-samples tests with Holm correction, Mann–Whitney tests, psychometric assessment, heteroscedasticity-robust regressions, ordered and binary logistic models, propensity-score overlap weighting, heterogeneity tests, and temporal sensitivity analysis. The largest difference concerned argument persuasiveness: the mean score was 3.56 under ChatGPT versus 2.66 for static media, yielding a 0.90-point difference, Hedges’ g of 0.823, and p < 0.001, while the adjusted coefficient reached 1.122 with p < 0.001. Perceived understanding was also higher in the ChatGPT condition, at 4.04 versus 3.52, with a 0.52-point difference, Hedges’ g of 0.520, and a Holm-adjusted p-value of 0.031; the adjusted high-response probability was 89.9% for ChatGPT versus 56.0% for static media. Perspective influence showed a smaller unadjusted difference of 0.42 points, but the heteroscedasticity-robust adjusted coefficient reached 0.634 with p = 0.002, and the adjusted high-response probability was 68.6% for ChatGPT versus 38.6% for static media, whereas explicit opinion change remained the least robust outcome, and the overall post-exposure composite was 0.500 points higher under ChatGPT with p = 0.0045 and Hedges’ g of 0.578. These findings provide a basis for future research and practical guidance on responsible artificial-intelligence-mediated communication by demonstrating differentiated associations with understanding, perceived persuasiveness, and reported opinion-related responses, while highlighting transparency, source credibility, argument quality, and user autonomy as important considerations for the responsible deployment of conversational systems.