I recently gave a talk at the AI Research Centre seminar series at City St George’s, University of London, titled “When Good Metrics Lie: What the Listener Acoustic Personalisation Challenge Revealed About AI in Spatial Audio.” The talk explored what the Listener Acoustic Personalisation Challenge 2024 (LAP24) revealed about the limits of standard objective metrics in spatial audio, and why good numerical scores do not always translate into perceptually meaningful personalisation. More broadly, it reflected on what this tells us about evaluating AI systems that aim to model human perception.