What We Learned Trying to Catch AI Liars: An Aletheia's Quest Retrospective

A team from EleutherAI spent four weeks competing in Aletheia's Quest, a July 2026 competition where 19 teams raced to build the best AI lie detector, with $50,000 in prizes on the line. Their two-phase detection method used a trusted judge to independently answer questions and then check AI responses for false claims, achieving 0.926 AUROC on the black-box leaderboard. They led for most of the competition before being edged out in the final days by SAIN Groningen, whose single-pass approach was faster and slightly more accurate.