Trust and Usefulness of Extractive Rationales: A Comparison Between Human-generated and LLM-generated Explanations.
Keywords
Loading...
Authors
Issue Date
2025-06-23
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
The present study investigated differences between LLM-generated and human-generated explanations in the form of extractive rationales (a set of words extracted from input texts that illustrate the decision-making process) for the disaster tweet classification task. The differences in users’ perceived trust and usefulness were examined in a between-subject experimental study with a total of 211 participants.
Results of the study showed no significant differences between the two conditions for trust and distrust dimensions. For both sources of outputs, alignment of rationale and tweet classification label was seen as fairly good, while usefulness of rationale was seen as medium. Findings shed a light on the user’s perceived attitude towards rationales, showing that when focused on the output itself, the source of it might not be of high influence. However, further research is encouraged, as extractive rationales remain quite a new concept for the general public yet hold immense potential for promoting transparent AI tools.
Description
Citation
Supervisor
Faculty
Faculteit der Letteren