According to a study, ChatGPT makes many mistakes when answering programming questions. However, in many cases, users seem to be confident in the answer.

ChatGPT and other generative AI tools are touted as a revolution for programming. Developers just have to say what they want to develop and the AI ​​does the rest. In practice, it is a good idea to carefully review AI-generated code before running it, because ChatGPT doesn’t seem to be that great a programmer after all.

Researchers at Purdue University in the US examined 517 randomly selected questions answered using ChatGPT on the developer forum Stack Overflow. 52 percent of these answers contained incorrect information, while 78 percent of the answers did not match human answers to similar questions. ChatGPT also uses significantly more words than human developers.

Friendly tone

The length of the answers misleads people. As part of the study, twelve developers also read answers from AI and humans and chose the answer they liked best. In 35 percent of cases, the AI-generated answer was chosen.

The researchers cite the length and detail of the answers as well as the friendly tone and eloquent writing style as the main reasons for this. As a result, the test subjects often overlooked incorrect information. The experiment and the results are described in detail in a research paper.

This is of course just one study, but it shows that we need to remain critical of AI tools and not blindly trust their word. AI makes mistakes, just like humans, and ChatGPT is no exception. It could always be worse: Google’s new “AI search engine” produces utter nonsense.