Abstract
AI demonstrates unprecedented reasoning capabilities, but its increasing integration into human reasoning via automated reading and summarization has provoked debate about its use for cultural interpretation. Close reading—the practice of understanding, analyzing, and critiquing cultural texts for pleasure—is a skill at the core of such interpretation, traditionally being seen as exclusive to humans. To test AI's impact on close reading, both in terms of interpretative performance and pleasure, we conducted a preregistered randomized experiment ($n=400$) investigating the impact of AI assistance by presenting single or multiple AI interpretations, on close reading poems, compared to no AI assistance. We found that single AI interpretation boosted both performance and pleasure, while multiple AI interpretations only improved performance. Further exploration revealed a trade-off: participants who heavily relied on AI showed better performance on the task but lower pleasure. Our results contribute to discussion on whether and how to calibrate AI assistance for cultural interpretation: "less is more."
Introduction
As Large Language Models (LLMs) become integrated into reading platforms, they are increasingly used to summarize and interpret cultural texts like poems, stories, and films. While AI is excellent at "cognitive outsourcing"—doing the heavy lifting of summarization—cultural interpretation (or "close reading") is unique. Its value lies not just in the final "answer," but in the process of discovery, the "aha" moment, and the personal resonance a reader feels. This study, presented at CHI 2026, explores the tension between AI’s analytical power and the human experience of art, identifying a "less is more" principle for AI-assisted reading.
Methods
The researchers conducted a preregistered, between-subjects randomized controlled trial with 400 participants on Prolific. Participants were asked to perform "close reading" tasks on three curated poems: “Love Poem” by Linda Pastan, “Dusting” by Marilyn Nelson, and “Theme for English B” by Langston Hughes. The task involved identifying stylistic features (like metaphors or rhythm) and explaining their effects.
Participants were randomly assigned to one of three conditions:
- Control: Read and interpret the poems with no AI assistance.
- AI-Single: Exposed to one AI-generated interpretation (using GPT-4o).
- AI-Multiple: Exposed to three distinct AI-generated interpretations.
The team measured two sets of dependent variables: Interpretive Performance (graded by human experts on feature identification, interpretation quality, and writing quality) and Subjective Experience (self-reported ratings of appreciation, enjoyment, and self-efficacy).
Key Results and Insights
The experiment revealed that while AI can significantly boost analytical performance, its impact on the reader's experience is more complex.
The Performance-Pleasure Trade-off
Both AI-Single and AI-Multiple conditions significantly improved interpretation and writing quality compared to the control group. However, only the AI-Single condition significantly increased enjoyment and appreciation. Participants in the AI-Single group were 2.64 times more likely to report higher enjoyment ratings than the control group. In the AI-Multiple condition, there was no significant increase in pleasure measures.
The Threat of Competence
For experienced readers (those with college-level humanities coursework), multiple AI interpretations led to a significant drop in self-efficacy ($p = 0.021, r = -0.257$). These readers felt the AI had "exhausted the interpretive space," leaving them with nothing original to contribute. This suggests that "maximal" AI assistance can be demotivating for those who value their own interpretive skills.
Ambient Learning and Heavy Reliance
Interestingly, even participants in the AI-Single group who claimed they didn't use the AI still performed better than the control group, suggesting an "ambient learning" effect where high-quality examples help readers model better strategies. Conversely, there was a clear "heavy copier" penalty: participants who relied most heavily on AI (high textual overlap) achieved the highest performance scores but reported the lowest levels of enjoyment and appreciation.
Conclusion
The study concludes that a single AI interpretation acts as a helpful "scaffolding" or guide, providing a quality benchmark without overwhelming the reader. However, too much AI assistance can drain the pleasure and self-efficacy out of the reading process. Future research should investigate interactive AI models and the timing of assistance to better preserve human autonomy and pleasure in cultural interpretation.
Summary updated