Advanced AI Chatbots Hesitant to Admit Their Limitations

The study also found that people often too readily accept incorrect answers from bots. Do you think this trust in bots poses a danger to us?

Advanced Chatbots: A Trend Towards Misinformation

Artificial intelligence advancements have led to increasingly sophisticated chatbots. However, a recent study highlights a concerning trend: these advanced AIs are more prone to providing incorrect answers rather than admitting their lack of knowledge.

A Study on Three Language Model Families

Led by José Hernández-Orallo, a professor at the Universitat Politecnica de Valencia, the study focused on three automatic language model families: OpenAI’s GPT, Meta’s LLaMA, and the open-source BLOOM. The researchers tested earlier and more advanced versions of each model, excluding the most recent ones.

The models were challenged with thousands of questions across various fields such as “arithmetic, anagrams, geography, and science”. The team also assessed their ability to process information, like alphabetically sorting a list.

Increasing Incorrect Responses

The findings indicated that the proportion of incorrect responses from chatbots increased as the models developed. It’s akin to a teacher, gaining mastery over more subjects, starting to believe they have all the answers.

Furthermore, volunteers evaluating the accuracy of the AI responses often mistakenly rated inaccurate answers as accurate. The error rate perceived as correct generally varied between 10 and 40%.

Recommendations to Enhance Chatbot Reliability

In light of these findings, researchers recommend focusing on improving chatbot performance for simple questions and programming them to refrain from answering more complex ones. They believe users need to understand which areas these tools are suitable for and which they are not.

It is important to note that these recommendations might be hard for AI companies to accept, as a chatbot frequently saying “I don’t know” could be seen as less advanced or less valuable, potentially leading to decreased usage and therefore lower revenues for the companies that produce and sell them.

In the meantime, it is up to us, the users, to verify the accuracy of our chatbots’ responses to prevent believing and spreading misinformation that could harm us or others.