Narrative Transportation and Speciesism in Large Language Models

In humans, narratives often influence beliefs more strongly than equivalent fac- tual descriptions, an effect known as narrative transportation. Whether narratives similarly influence the attitudes expressed by language models remains unknown. Following the design of Green and Brock’s narrative transportation experiments, we had five instruction-tuned language models read a story about an animal, a matched factual version, or nothing, and then answer a speciesism questionnaire, for each of 12 animals. Narratives reduced expressed speciesism on the questions about the animal they described, whereas the matched factual versions generally did not, and neither did an explicit essay arguing for animal consider- ation, suggesting that the effect is specific to narrative framing rather than informational content. The models’ own transportation ratings tracked how immersive the writing was, but the reduction did not clearly grow with rated transportation, so we make no claim about a transportation-like mechanism. These findings suggest that narratives can influence the positions language models express, and that such effects are easiest to detect at the level of individual questionnaire items.