Persuasive messages written by generative AI are easier to read, liked better, and perceived as more probably true than messages written by humans

Prior research has shown that Large Language Models (LLMs) can produce effective persuasive messages. One reason why LLM-generated messages are persuasive could be that they are easy to read, which could enhance their credibility and likability. This paper examines the reading ease, perceived truth, and likability of persuasive messages produced by GPT-4o compared to messages authored by human lay writers. Although human-written messages were objectively easier to read, Experiments 1 and 2 showed that readers perceived LLM-generated persuasive messages as easier to read, as more truthful, and liked these more than human-written messages. Experiment 3 showed that this pattern was replicated when we matched the objective reading ease of LLM-generated and of human-written messages. Experiment 4 replicated these results when we additionally matched the arguments of LLM-generated and of human-written messages thereby controlling for the prototypicality of the arguments. Our findings suggest that the ease with which human readers can understand text generated by LLMs could contribute to its persuasiveness. We discuss these results in context of a predictive coding account of language comprehension according to which humans constantly predict the continuation of a text while reading, which resembles the way LLMs generate text.