Good Answers Are Not Necessarily Factual Answers

Phare’s new benchmark reveals that leading LLMs confidently produce fabricated information despite sounding authoritative. Their evaluation of models from eight AI labs shows these systems regularly generate completely fabricated details when handling misinformation queries. The benchmark tests models on factuality, misinformation resistance, debunking capabilities, and tool reliability. Three key findings stand out: Popular isn’t truthful: Models ranking highest in user satisfaction...