A study by UNESCO and University College London warns against relying on AI chatbots to guide climate policy decisions. Tests involving up to 33 models showed that they correctly answered around 90% of neutral climate questions in English. However, when users challenged their answers, eight out of ten chatbots changed their position, sometimes even agreeing with false claims.
The models also struggled to assess new evidence and understand how different climate measures interact. Performance was even weaker in several other languages. The study criticises AI systems for prioritising user agreement and satisfaction over facts and accuracy. Recommendations include targeted training, independent fact-checking, regular answer verification, improved multilingual capabilities and human expert oversight for complex decisions.