One of the big selling points of using generative AI chatbots like Copilot, Gemini, and ChatGPT is that they can help save time by answering questions about coding and programming faster than what could be done with humans answering those same prompts. However, a recent report claims that ChatGPT is offering a lot of incorrect answers to programming questions.
In a report that was presented earlier in May during the Computer-Human Interaction Conference (via Gizmodo), a research team from Purdue University checked out 517 Stack Overflow questions that were sent into ChatGPT to answer.
The Purdue University team reported back that ChatGPT offered incorrect answers to those programming questions a whopping 52 percent of the time.
While offering incorrect answers to questions is perhaps somewhat expected from an AI chatbot, the Perdue University team went further, stating that human programmers still preferred the ChatGPT answers 35 percent of the time "due to their comprehensiveness and well-articulated language style." What's worse, the human programmers didn't catch the answers ChatGPT made in error 39 percent of the time.
While this is just one study of one aspect of ChatGPT's features, it does show that generative AI bots are still prone to making lots of mistakes and that humans may not catch those mistakes.
Google's AI Overviews, which was put in place in the US region for Google Search earlier in May, have already been generating strange and error-filled answer summaries to some search strings. Google has tried to explain away these obvious errors in statements made to news outlets like Gizmodo, stating:
The examples we’ve seen are generally very uncommon queries, and aren’t representative of most people’s experiences. The vast majority of AI Overviews provide high quality information, with links to dig deeper on the web.
The statement added Google would use these "isolated examples" to help "refine our systems."
14 Comments - Add comment