Testing the Reliability of ChatGPT for Text Annotation and Classification: A Cautionary Remark
Michael V. Reiss
TL;DR
Results show that consistency in ChatGPT's classification output can fall short of scientific thresholds for reliability, and advises caution when usingChatGPT for zero-shot text annotation and underscores the need for thorough validation, such as comparison against human-annotated data.
Abstract
Recent studies have demonstrated promising potential of ChatGPT for various text annotation and classification tasks. However, ChatGPT is non-deterministic which means that, as with human coders, identical input can lead to different outputs. Given this, it seems appropriate to test the reliability of ChatGPT. Therefore, this study investigates the consistency of ChatGPT's zero-shot capabilities for text annotation and classification, focusing on different model parameters, prompt variations, and repetitions of identical inputs. Based on the real-world classification task of differentiating website texts into news and not news, results show that consistency in ChatGPT's classification output can fall short of scientific thresholds for reliability. For example, even minor wording alterations in prompts or repeating the identical input can lead to varying outputs. Although pooling outputs from multiple repetitions can improve reliability, this study advises caution when using ChatGPT for zero-shot text annotation and underscores the need for thorough validation, such as comparison against human-annotated data. The unsupervised application of ChatGPT for text annotation and classification is not recommended.