Advanced AI chatbots are less likely to admit they don’t have all the answers

Researchers have seen the downside of intelligent chatbots. Although AI predictive models are becoming more accurate as they improve, they are also more likely to (erroneously) answer questions beyond their capabilities than, “I don’t know.” And the people who encourage them are more likely to take their misinformation into account, creating a diminishing effect of misinformation.

“They answer almost everything these days,” José Hernández-Orallo, a professor at the Universitat Politecnica de Valencia, Spain, told. The environment. “And that means it’s right, but also very wrong.” Hernández-Orallo, the project leader, worked on the research with colleagues at the Valencian Research Institute for Artificial Intelligence in Spain.

The team studied three LLM families, including OpenAI’s GPT series, Meta’s LLaMA and the open-source BLOOM. They tested the earlier versions of each model and moved to the larger, more advanced – but not the most advanced of today. For example, the team started with OpenAI’s ada model of the old GPT-3 and tested iterations leading to GPT-4, which arrived in March 2023. The four-month-old GPT-4o was not included in the study, and was not included. new o1 preview. I would like to know if the trend is still with the latest models.

The researchers tested each model on thousands of questions about “math, anagrams, geography and science.” They also questioned the AI ​​models about their ability to manipulate information, such as alphabetizing lists. The team ranked their information by perceived difficulty.

The data showed that the proportion of incorrect answers by chatbots (instead of avoiding questions entirely) rose as the models grew. Therefore, AI is like a professor who, as he studies more subjects, increasingly believes that he has the golden answers to all of them.

Another irony is that people encourage chatbots and read their responses. The researchers tasked volunteers with rating the accuracy of AI bots’ answers, and found that they “misclassified incorrect answers as surprisingly accurate.” The range of incorrect answers that the volunteers incorrectly perceived as correct generally fell between 10 and 40 percent.

“People cannot direct these examples,” concluded Hernández-Orallo.

The research team recommends that AI developers start increasing the performance of simple questions and program chatbots to refuse to answer complex questions. “We need people to understand: ‘I can use it in this place, and I shouldn’t use it in that place,'” Hernández-Orallo told. The environment.

It’s a well-intentioned proposition that would make sense in an ideal world. But AI companies’ fat opportunities are compelling. Chatbots that often say “I don’t know” may be considered low-level or irrelevant, leading to less use – and less money for the companies that make and sell them. So, instead, we get the fine-printed warnings that “ChatGPT may make errors” and “Gemini may display incorrect information.”

That leaves it in our hands to avoid believing and spreading false information that could harm us or others. For accuracy, check your chatbot’s responses, loud and clear.

You can read the group’s full tutorial at The environment.