Asking chatbots for short answers can increase hallucinations, study finds

It turns out that an AI chatbot can be shorter than the way it can be otherwise hallucinet.

It is developing an overall standard for AI models according to a new survey by the Paris-based AI testing company Giskard. A Blog post Giskard’s researchers say that their search details say that the questions specially for the short answers to the question, especially the questions about vague things, can have a negative impact on the reality of an AI model.

Researchers wrote, “Our data shows that simple changes in the system’s instructions dramatically affect the tendency of a model to hallucinet,” researchers wrote. “There is an important impact for deploying for this search, as many applications give priority to reduce short outputs [data] Use, improve the delay and reduce the cost. “

Hallucination A complicated problem in AI. Even the most capable models sometimes make things, a feature of them Potential NatureThe In fact, new rational models like Openai and 3 Hallucinate More It makes their results hard to trust than previous models.

In its research, Giskard identified certain requests that could make hallucinations worse, such as obscure and incorrect information asking for short answers (eg, “tell me why Japan won the WII”). OpenAI’s GPT -4 (default model Powering ChatzPT), Mistral Large and ethnic Claud 3.7 Sonnets are asked to keep answers short when asked to keep answers short.

Why? GISCARD assumed that when asked for not to answer greatly, models are not “space” to just recognize false premises and indicate mistakes. In other words, strong rejections require a long explanation.

Researchers wrote, “When it is forced to keep it short, the models are consistently choosing Bravy rather than accuracy.” “Perhaps for the most important developers, the seemingly innocent system requests like ‘B Connis’ as promise that a model can destroy the ability to debt incorrect information.”

There are other curious revelations in the study of Giskard, as the models are less likely to reduce controversial claims when users are confidently presenting and the models that users say they love are not always the most truthful. In fact, there is OpenAI Recently struggled Maintaining a balance between models that give validity without having to come as excessively psychopantic.

Researchers wrote, “Optimization for user experience can sometimes come to the cost of true accuracy.” “It creates a tension between accuracy and alignment with user expectations, especially when those expectations include false premises.”