ChatGPT: Scientists warn programmers about errors

To speed up the workflow, many programmers now rely on the support of ChatGPT. However, a recently published study by researchers now shows that this is probably not a good idea. Some of the AI results are said to contain serious errors.

Warning about ChatGPT in IT

In IT, ChatGPT has proven to be a real blessing for many. After all, programmers like to use the chatbot not only to have their source code checked. Some even have entire codes written by the AI. Now a team of scientists has presented a study at CHI 2024 that makes it clear that using them is probably not such a good idea. As part of the study, the researchers took a close look at several chatbots.

However, the results of the AI-supported little helpers were very similar. The study revealed that the answers from ChatGPT and co. contain various errors. A whopping 50% of the results produced by the AI were riddled with errors. In their study, the scientists used a pool of questions from the Stack Overflow platform. The questions posed there were answered accurately by ChatGPT.

On closer inspection, however, the researchers noticed that the codes issued by the AI contained various errors. An article by our colleagues at Golem.de reveals that the AI apparently has problems grasping context. Unsurprisingly, the problems were mainly with questions that were rarely asked or answered. After all, the chatbot likes to use such forums as training data.

Faulty, but friendly

Despite the errors, many programmers are enthusiastic about their work with the support of ChatGPT and Co. This is apparently due to the fact that the AI’s answers are not only formulated in detail. On top of that, the tools are careful to use as friendly a tone as possible. This is not usually the rule in forums that are used to solve problems.

Nevertheless, the researchers now see a reason for the IT industry to react. Users should not blindly seek the support of ChatGPT and co. Rather, it is important to thoroughly check the results of the AI once again. Otherwise, serious software errors could result, which would certainly not be in the interests of the programmers.