Hallucination or Creativity: How to Evaluate AI-Generated Scientific Stories?

TL;DR

The paper tackles evaluating AI-generated scientific stories that adapt to audience personas, proposing StoryScore as a composite metric balancing semantic fidelity, lexical grounding, structural fidelity, fluency, and hallucination control. It introduces a concrete use case with a two-stage generation pipeline and analyzes several hallucination detectors, finding that traditional similarity metrics miss narrative-level issues and that simple entity-based detectors offer more stability than complex LLM judges or MIRAGE-style methods. Empirical results show that fine-tuning improves StoryScore and contextual grounding, while prompt leakage and repetition remain challenges; hallucination detection remains intrinsically difficult when creativity and pedagogical reformulation are desired. The work highlights the need for multi-signal evaluation in scientific storytelling and outlines directions for future development, including explicit assessment of narrative control and alignment with human judgment.

Abstract

Generative AI can turn scientific articles into narratives for diverse audiences, but evaluating these stories remains challenging. Storytelling demands abstraction, simplification, and pedagogical creativity-qualities that are not often well-captured by standard summarization metrics. Meanwhile, factual hallucinations are critical in scientific contexts, yet, detectors often misclassify legitimate narrative reformulations or prove unstable when creativity is involved. In this work, we propose StoryScore, a composite metric for evaluating AI-generated scientific stories. StoryScore integrates semantic alignment, lexical grounding, narrative control, structural fidelity, redundancy avoidance, and entity-level hallucination detection into a unified framework. Our analysis also reveals why many hallucination detection methods fail to distinguish pedagogical creativity from factual errors, highlighting a key limitation: while automatic metrics can effectively assess semantic similarity with original content, they struggle to evaluate how it is narrated and controlled.