Say Goodbye to Late Parkinson’s Diagnosis with New AI Blood Test!
June 20, 2024
Cyber Babe Empowering Communities: Moana Tearei’s Inspiring Journey as an Animal Shelter Volunteer on Cook Islands!
June 20, 2024
Scientists have a new method to ensure AI doesn’t ‘hallucinate’. Find out how!
Tackling AI Hallucinations: New Research Offers Hope
Generative AI tools like ChatGPT often produce false information, a phenomenon known as “hallucination.” These errors have led to notable public mishaps, such as AirCanada having to honour a mistaken discount and Google modifying its AI search feature after incorrect suggestions. Recently, two lawyers were fined for using ChatGPT-generated citations that referenced nonexistent cases. However, a study published in Nature introduces a method to detect AI hallucinations with approximately 79% accuracy, surpassing current methods by about 10 percentage points. This technique could make AI tools more reliable, although it requires significantly more computing power.
Understanding and Detecting AI Confabulations
Sebastian Farquhar, a senior research fellow at Oxford University and a research scientist at Google DeepMind, led the study, focusing on a specific type of AI error called “confabulations.” Unlike consistent errors, confabulations occur when AI gives different incorrect answers to the same question. The researchers’ method involves generating multiple answers to a prompt, clustering them by meaning, and calculating “semantic entropy” to measure answer consistency. High entropy indicates likely confabulations, while low entropy suggests consistency, even if the answer is wrong. This method outperformed several others, including naive entropy and embedding regression.
Implications and Future Integration
Farquhar envisions practical applications, such as a certainty score feature for AI-generated answers, enhancing user confidence in AI tools. However, integrating this method into real-world applications poses challenges. Arvind Narayanan, a Princeton University computer science professor, appreciates the research but warns against overestimating its immediate impact. He notes that while hallucination rates have decreased with better models, the issue may persist due to the intrinsic nature of large language models. As AI’s capabilities grow, so do the complexity of tasks they are used for, making the elimination of hallucinations a difficult, ongoing challenge.
(Visit Time for the full story)
*An AI tool was used to add an extra layer to the editing process for this story.
