Evaluating the Influence of Role-Playing Prompts on ChatGPT's Misinformation Detection Accuracy: Quantitative Study

Area of Science:

• Artificial Intelligence
• Public Health Communication
• Computational Social Science

Background:

• The COVID-19 pandemic highlighted challenges in combating online misinformation.
• Large language models (LLMs) show promise for misinformation detection but are sensitive to prompt engineering.
• Role-playing prompts, where LLMs adopt specific identities, may influence their performance.

Purpose of the Study:

• To assess how assigning social identities to ChatGPT via role-playing prompts affects its accuracy in detecting COVID-19 misinformation.
• To compare ChatGPT's performance on explicit versus context-dependent misinformation.
• To analyze the reasoning provided by ChatGPT for its classification decisions.

Main Methods:

• 36 COVID-19 related tweets were classified (misinformation, sentiment, corrections, neutral).
• ChatGPT was tested with prompts including various combinations of social identities (political, educational, locality, religiosity, personality).
• Control conditions included prompts with no identities and only political identity, totaling 51,840 runs.

Main Results:

• Including social identities in prompts decreased average misinformation detection accuracy from 68.1% to 29.3%.
• Prompts with only political identity yielded the lowest accuracy (19.2%).
• ChatGPT distinguished between non-aligned sentiment and misinformation, but reasoning was inconsistent and explanations contradictory.

Conclusions:

• Role-playing social identities negatively impacts ChatGPT's misinformation detection capabilities.
• The integration of human biases into LLMs is complex, necessitating human oversight.
• Further research is required to understand LLM processing of social identities and cross-cultural applications.