What Makes AI a Good Cultural Mediator? Evidence from Literary Paratexts

Zhou Mengyuan

Keywords: cultural AI, AI evaluation, cultural mediation, interpretive technologies, literary paratexts, cultural grounding, large language models

TL;DR: A framework for evaluating AI as a cultural mediator, tested on AI-generated literary paratexts.

Abstract: Generative AI systems are increasingly used to introduce, explain, recommend, and reframe cultural works for new audiences, yet cultural AI evaluation is often framed either as harm detection or as generic text generation quality. We propose a framework for evaluating AI as a cultural mediator: a system that helps audiences encounter cultural objects through generated descriptions, explanations, or recommendations. The framework defines four operational dimensions of good cultural mediation: availability, interpretive substance, cultural grounding, and discursive alignment. We demonstrate the framework on AI-generated literary paratexts using 282 contemporary literary titles, 28,200 outputs, and five model families. Original-language and English generation are used as diagnostic conditions. The framework reveals a trade-off that would be difficult to observe with a single score: English prompting reduces refusals from 64 to 24 and generic failures from 43 to 19, but also reduces average entity recall from 0.1819 to 0.1135 and average Jaccard overlap from 0.1169 to 0.0734. The paper contributes a task-grounded evaluation framework that makes visible when gains in usability coincide with losses in cultural grounding.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 3