Advanced AI chatbots reluctant to admit their limits

The study also found that people tend to be too quick to accept bots’ wrong answers. Do you think this gullibility towards bots is a danger to us?

Advanced Chatbots: A Trend of Disinformation

Advances in artificial intelligence have made it possible to create increasingly powerful chatbots. However, a recent study highlights a worrying trend: these advanced AIs are increasingly likely to give incorrect answers rather than admit their ignorance.

A study of three families of language models

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

The models were asked thousands of questions covering a variety of topics, including “arithmetic, anagrams, geography and science.” The team also tested their ability to transform information, such as alphabetizing a list.

Wrong answers on the rise

The data collected showed that the proportion of incorrect answers from chatbots increased as the models evolved. It’s a bit like a teacher, as he masters more and more subjects, starting to believe he has all the answers.

Additionally, volunteers tasked with assessing the accuracy of AI responses often misclassified inaccurate answers as correct. The rate of errors falsely perceived as correct typically ranged from 10 to 40 percent.

Recommendations to improve the reliability of chatbots

Given these findings, the researchers recommend focusing efforts on improving the performance of chatbots for simple questions and programming them to refuse to answer more complex questions. They believe that users need to understand in which areas they can use these tools and in which they should not.

It is important to note that these recommendations might be difficult to accept for AI companies, because a chatbot that often says “I don’t know” might be perceived as less advanced or less valuable, which could lead to a decrease in its use and therefore revenue 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 avoid believing and spreading misinformation that could harm us or others.

Shares:
Leave a Reply

Your email address will not be published. Required fields are marked *