# A Categorical Archive of ChatGPT Failures

**Authors:** Ali Borji  
**arXiv:** [2302.03494](https://arxiv.org/abs/2302.03494)  
**Date:** 2023-02-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.03494v1/

## Key points

- Large language models (LLMs) like ChatGPT have proven valuable in various fields
- ChatGPT can simulate human conversation and generate appropriate responses
- ChatGPT's ability to answer a broad range of human inquiries surpasses prior public chatbots in both security and usefulness
- A comprehensive analysis of ChatGPT's failures is lacking
- Ten categories of failures exhibited by ChatGPT are presented, including reasoning, factual errors, math, coding, and bias
- The risks, limitations, and societal implications of ChatGPT are highlighted
- LLMs like ChatGPT have limitations and can often generate incorrect information
- Efforts are ongoing to formulate standardized sets of questions to track the progress made by these models over time
- It is necessary for researchers and developers to address these limitations while enhancing future language models and chatbots' development
- This will maximize their potentials while minimizing risks associated with them such as bias or incorrect information generation.

## Summary

Large language models (LLMs) have proven to be valuable in various fields, and ChatGPT, developed by OpenAI, is a prime example of this. It has been trained using massive amounts of data and can simulate human conversation by comprehending context and generating appropriate responses. ChatGPT's ability to answer a broad range of human inquiries with fluent and comprehensive answers surpasses prior public chatbots in both security and usefulness. However, despite its impressive capabilities, a comprehensive analysis of ChatGPT's failures is lacking. This study aims to fill that gap by presenting ten categories of failures exhibited by ChatGPT, including reasoning, factual errors, math, coding, and bias. The risks, limitations, and societal implications of ChatGPT are also highlighted. While LLMs like ChatGPT have proven useful in several areas such as conversational agents, education, explainable AI, text summarization, information retrieval among others; they are not without their limitations and can often generate incorrect information. To fully leverage their capabilities it is crucial to acknowledge their limitations and biases in their generated output. To accurately assess the performance of these models over time instead of relying on subjective opinions; some efforts are ongoing in formulating standardized sets of questions which will track the progress made by these models over time. Ernest Davis has proposed such question sets for testing the knowledge base used by large language models like GPT-3. While future studies may enhance our understanding of LLMs' capabilities further; it is essential to recognize that these models are not infallible. As outlined by Ali Borji's GitHub repository on ChatGPT failures; it is necessary for researchers and developers to address these limitations while enhancing future language models and chatbots' development in order to maximize their potentials while minimizing risks associated with them such as bias or incorrect information generation.