Regular users of ChatGPT are well-acquainted with its disclaimer, ‘ChatGPT can make mistakes. Check important info,’ and are often reminded of it. Computer programmers who use the service to help them code will want to take this warning more seriously, as a study has found that the chatbot gives the wrong answers more than half the time.
A research team from Purdue University presented a paper at an annual academic conference by the Association for Computing Machinery, examining how ChatGPT’s answers fare in quality and accuracy compared to those on StackOverflow. Based on 517 programming questions sourced from the Q&A platform, the researchers found that 52 percent of ChatGPT’s answers have misinformation in them.
The researchers also observed that ChatGPT provided verbose answers 77 percent of the time, meaning the chatbot often padded its responses with information deemed redundant, irrelevant and unnecessary. Additionally, the analysis found that ChatGPT’s answers were inconsistent with human answers on StackOverflow about 78 percent of the time.
- Learn programming and software development with these mobile apps
- Undetectable AI? OpenAI says AI writing detectors don’t work
Despite ChatGPT’s tendency to give incorrect, misleading, and inconsistent answers, the study found that its user participants still preferred the chatbot’s answers 35 percent of the time. The reason? ChatGPT gave comprehensive and well-articulated answers. This level of detailed response seemed to be convincing enough that users overlooked the misinformation in the answers 39 percent of the time.
Given the inaccuracies and verbosity in ChatGPT’s answers, the researchers recommend further research on countering misinformation and mitigating risks of errors in large-language models.