Study Highlights Growing Issue with AI Chatbots and Misinformation
Valencia, Spain — Researchers at the Valencian Research Institute for Artificial Intelligence have identified a significant challenge with the increasing sophistication of AI chatbots. Their study reveals that as AI models become more advanced, they are more likely to provide incorrect answers rather than admit their limitations, a phenomenon that could lead to widespread misinformation.
The research, headed by Professor José Hernández-Orallo from the Universitat Politecnica de Valencia, examined three prominent families of large language models (LLMs): OpenAI’s GPT series, Meta’s LLaMA, and the open-source BLOOM. The team analysed various iterations of these models, starting from relatively earlier versions, such as OpenAI’s GPT-3 ada model, and progressing toward more advanced iterations up to GPT-4, though the very latest models were not included in the study.
The researchers assessed the models by posing thousands of questions covering topics such as arithmetic, anagrams, geography, and science. They also tested the AI's ability to transform information, such as alphabetising lists, with complexities graded from easy to difficult.
Findings from the study show a troubling trend: as AI models become more advanced, their propensity to give wrong answers increases. This behaviour suggests that, similar to a professor who overestimates his expertise across various subjects, advanced AI is less likely to acknowledge its limitations. Instead, these models increasingly attempt to answer questions outside their capability with unwarranted confidence.
This issue is compounded by human interaction with these chatbots. The study involved volunteers who were asked to rate the accuracy of the AI-generated answers. Alarmingly, these volunteers often misjudged inaccurate responses as correct. The proportion of wrongly perceived correct answers ranged from 10 to 40 percent, underscoring humans' difficulty in supervising and verifying these models.
“Humans are not able to supervise these models,” concluded Hernández-Orallo.
The study underscores a need for AI developers to enhance performance accuracy for simple questions and program the chatbots to acknowledge their limitations more transparently. “We need humans to understand: ‘I can use it in this area, and I shouldn’t use it in that area,’” Hernández-Orallo stated.
However, practical implementation of these improvements faces challenges. Chatbots designed to more frequently admit to not knowing an answer might be seen as less advanced or valuable, potentially reducing their usage and profitability, something that global AI companies might be reluctant to adopt.
As it stands, the current AI chatbots include disclaimers, such as “ChatGPT can make mistakes” and “Gemini may display inaccurate info,” but the onus remains on users to be vigilant against accepting and spreading potential misinformation.
The full findings of this research are available in the journal Nature.
Source: Noah Wire Services