# Dr ChatGPT, tell me what I want to hear: How prompt knowledge impacts health answer correctness

**Authors:** Guido Zuccon, Bevan Koopman  
**arXiv:** [2302.13793](https://arxiv.org/abs/2302.13793)  
**Date:** 2023-02-23  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2302.13793v1/

## Key points

- The study explores the impact of prompt knowledge on the correctness of answers generated by generative pre-trained language models (GPLMs) like ChatGPT in the context of consumers seeking health advice.
- Prompt knowledge can override the model's encoded knowledge, leading to a decrease in answer correctness.
- The effectiveness of ChatGPT in answering health-related questions is demonstrated with an accuracy rate of 80%.
- Prompt knowledge often overturns model-generated answers about treatments, resulting in a decrease in overall accuracy (63%).
- Incorporating prompt knowledge can impact the reliability of GPLMs in providing accurate health advice.
- Further development is needed to ensure trustworthy outcomes.

## Summary

This paper titled "Dr ChatGPT, tell me what I want to hear: How prompt knowledge impacts health answer correctness" explores the impact of prompt knowledge on the correctness of answers generated by generative pre-trained language models (GPLMs) like ChatGPT in the context of consumers seeking health advice. GPLMs encode knowledge observed during pre-training and utilize it during inference to address user prompts. Additionally, they can leverage knowledge provided in the prompts themselves. The study aims to compare answer correctness when ChatGPT relies solely on its internal knowledge versus when it combines that knowledge with prompt knowledge. The researchers find that prompt knowledge can override the model's encoded knowledge, leading to a decrease in answer correctness. This work has significant implications for developing more robust and transparent question-answering systems based on GPLMs. The paper also investigates the effectiveness of ChatGPT in answering health-related questions and demonstrates its accuracy rate of 80%. Furthermore, when prompting with supporting or contrary evidence, they observe that prompt knowledge often overturns model-generated answers about treatments, resulting in a decrease in overall accuracy (63%). Overall, this research sheds light on how incorporating prompt knowledge can impact the reliability of GPLMs in providing accurate health advice and emphasizes the need for further development in this area to ensure trustworthy outcomes.