A team from the University of Oxford has developed a statistical model that can detect when large language models (LLMs) used in generative AI chatbots are likely to produce incorrect answers. This phenomenon, known as hallucination, occurs when AI fabricates information in response to a query.
Their findings, published in the journal Nature, address a major concern in the use of generative AI, where advanced models, like those powering ChatGPT, can present made-up information as fact. This issue is particularly pressing as more students and professionals turn to AI for research and assignment help.
Dr. Sebastian Farquhar, one of the study’s authors, explained that the new method differentiates between a model’s uncertainty about what to say and how to say it. Despite this progress, he acknowledged that further work is necessary to address systematic errors in AI outputs.
The research suggests that while semantic uncertainty can enhance reliability in specific cases, continuous improvements are needed to tackle AI’s systematic and confident mistakes.
Subscribe
Sign-up to receive our newsletter
More from Technology
Sir Tim Berners-Lee Announces Closure of World Wide Web Foundation
The World Wide Web Foundation is closing after 16 years of promoting an open and secure web, Sir Tim Berners-Lee has announced.
Groundbreaking AI Blood Test Predicts Male Infertility
Harnessing the power of AI, researchers have developed a revolutionary blood test that predicts male infertility with 74% accuracy.
Student Builds Nuclear Fusion Reactor, Earns Top Marks
A 17-year-old student has built a nuclear fusion reactor in his school, overcoming scientific challenges and concerns from the school.
3D Printing Capabilities Growing in Health and Construction
3D printing for health and construction is improving: breakthrough techniques are opening up new possibilities.
