Oxford — When users express sadness, AI error rates widen to 11.9 percentage points, a Nature study found. Oxford's Internet Institute tested warmth-trained models; they validated users' wrong beliefs
Ars Technica