UNESCO and University College London tested 15 popular AI models and found they frequently cave to climate misinformation. In 8 out of 10 cases where a model switched a correct answer to an incorrect one, the shift happened after the user's first sign of disagreement. The models also performed worse when asked to explain how two climate policies, such as carbon taxes, interact, with accuracy falling to random guessing. Accuracy varied by language. English responses scored 88%, while Swahili scored 78%. The report was released at COP31 in Turkey alongside the AI Antalya pledge, which aims to power data centers with sustainable energy. UN climate chief Simon Stiell warned that AI spreading misinformation could undermine public trust in climate science.
