The reliability and security of large language models (LLMs) have received increasing attention. Recent research has revealed potential flaws in LLM, such as duplication of harmful information and logical contradictions. These problems pose serious challenges to the application and development of LLM and require further research and improvement. This article will focus on a study on ChatGPT conducted by the University of Waterloo in Canada, which found that ChatGPT has repeated harmful misinformation and self-contradiction in answering questions, and provides an in-depth analysis of its causes and effects.
Recent research has found that large language models such as OpenAI’s ChatGPT often suffer from repeated harmful misinformation. Researchers at the University of Waterloo in Canada conducted a systematic test of ChatGPT's understanding capabilities and found that GPT-3 contradicted itself in its answers and repeated harmful misinformation. They used different survey templates and asked more than 1,200 different statements to discover the problem.The results of this study highlight the challenges faced by large language models in practical applications, and also provide an important reference for improving the reliability and security of LLM in the future. Further research should focus on how to reduce harmful information and logical errors in LLM output to ensure its safe and reliable application in various fields. It is hoped that future research can find more effective solutions and improve the quality and safety of LLM.