Trustworthy LLM for Current Event Sourcing, 10-R6573

Background

SwRI initiated this project to enhance the accuracy of answering questions about current events using advanced AI systems. Traditional AI systems often generate incorrect or outdated information on recent events, posing significant risks and legal concerns around misinformation. By focusing on responsibly gathering and utilizing the most current information from the web, the project aims to develop a more reliable and legally sound AI system capable of accurately addressing questions about current events.

Approach

The objective of this project is to create a more trustworthy AI system for answering questions about current events. To achieve this, the team explores methods for building trust in AI technology that utilizes web data, ensuring compliance with legal data collection practices, and develops a proof-of-concept system to demonstrate its reliability. Key areas of focus include mitigating the risk of misinformation by addressing the high-risk issue of misattribution (incorrect sourcing of information) and adhering to website terms of service during data collection. Using the National Institute of Standards and Technology AI Risk Management Framework, public web crawling tools, and verification mechanisms, the system emphasizes providing verifiable responses with multiple direct quotes from accurate sources. This ensures that human-in-the-loop users can quickly and easily check the validity of the information.

Accomplishments

The project successfully developed a more reliable AI system with enhanced capability in sourcing and verifying current event information. By concentrating on misattribution, the approach significantly improves the speed and accuracy of human-in-the-loop (HIL) verification compared to other models.

Testing reveals that the SwRI system not only reduces verification time but also maintained a high trustability score, demonstrating reliable information presentation. A notable accomplishment is the model's effectiveness in detecting and managing sources with conflicting information, thus ensuring response integrity. These advancements establish the SwRI system as a robust tool for providing trustworthy answers about current events.

A secondary phase of this effort is currently underway, leveraging this approach for use in alert management systems. This new phase aims to provide users with quick and accurate alerts about critical information from text feeds they subscribe to.