Asking chatbots for short answers can increase hallucinations, study finds

One says that an AI said that Chatbot would be short, it can otherwise be more alusive.

This is due to a new study from Giscard, a Paris-based AI test company that produces a single benchmark for AI models. One Blog Post Researchers in Giscard, who explain their findings in detail, especially questions on uncertain topics, especially the questions on uncertain issues, can negatively affect the Figative of the AI.

“Our data shows that simple changes in system instructions have a significant impact on the trend of a model of a model,” the researchers wrote. “This finding has important effects for placement, because many applications give preference to short results [data] To minimize the use, delay and costs. “

Hallucinations It’s a difficult problem in AI. Even the most skilled models sometimes make everything, a feature of these possibility nature. In fact, more new thinking models like Openai O3 set a management more More than previous models, it is difficult to trust their exits.

In his research, GISCARD identified certain instructions such as illegal and incorrect information (eg Japan’s WE WHII), Open Models Openai’s GPT-4O (Standard Model Amplifier Chatgpt) warns you to keep the answers in a short time.

Why? Giscard talks about not responding to a large detail, the models do not have a “void” to admit fake buildings and mark the mistakes. Strong refutation, other words require a longer explanation.

“When they had to keep in a short time, the models choose the accuracy of the accuracy,” said researchers. “Perhaps the most important thing as the most important thing as the seemingly innocent system ‘short’ can sabotage a model of a model.”

There are other interesting verses in the investigation of Giscard, these models provide users with confidence and the models that users want to do is not the most accurate. Indeed, there is Openai recently struggled to strike a balance between models confirming without coming overage.

“Optimization for user experience sometimes comes at the expense of the actual accuracy,” he wrote researchers. “This creates a tension between user expectations and creates a tension between alignment, especially when these expectations are entered in fake rooms.”