ChatGPT’s Programming Answers are Inaccurate Over 50%: Study

Study finds ChatGPT's programming answers unreliable in more than half of the cases

Purdue University research found that over half the time, OpenAI's fabulating chatbot, ChatGPT, answers questions on software development incorrectly. Despite this, a third of the participants were duped by the bot.

The Purdue team examined ChatGPT's Programming Answers to 517 Stack Overflow queries to evaluate the accuracy, coherence, breadth, and concision of ChatGPT's responses. Additionally, the US academics performed language and sentiment analysis on the responses before asking a dozen willing volunteers about the model's output.

The team's paper said, "Our analysis shows that 52% of ChatGPT answers are incorrect and 77% are verbose." However, owing to their thoroughness and well-articulated linguistic style, Chatgpt Answers are still favored 39.34% of the time. 77 percent of the chosen ChatGPT responses needed to be corrected.

On ChatGPT, OpenAI admits that its program "may produce inaccurate information about people, places, or facts." The lab has been questioned to see whether it has any comments on the Purdue study.

Who Answers It Better? is the title of the pre-print paper. An in-depth examination of ChatGPT and Stack Overflow responses to questions on software engineering. Researchers David Udo-Imeh, Bonan Kou, Samia Kabir, and Tianyi Zhang, an assistant professor, wrote it.

Their report found that users could only detect errors in ChatGPT answers when they were evident. Users frequently need to recognize the solution's incorrectness or underestimate its mistake level, however, when the fault is not easily verified or needs an external IDE or documentation.

The study found that two out of the 12 individuals continued to select the response as preferable even though it contained a clear inaccuracy. The friendly, authoritative tone of ChatGPT is credited with this, according to the report.

According to semi-structured interviews, "it is clear that polite language, articulated and textbook-style answers, comprehensiveness, and affiliation in answers make completely incorrect answers seem correct," the research said.