Researchers from University College London and UNESCO evaluated 33 large language models on their ability to provide accurate climate information. The study tested models using 600 questions across various benchmarks, including textbook facts and scientific report interpretations.

Frontier models achieved approximately 78 percent accuracy on complex scientific questions, while smaller open-weight models scored around 47 percent. On simpler benchmarks, top models reached 90 percent accuracy, whereas smaller models scored in the 70s.

The study examined model performance on policy impact simulations and evidence evaluation. While models answered single-policy questions with 72 to 89 percent accuracy, their performance on combined policy scenarios was no better than random guessing.

Testing revealed that models are susceptible to user pressure. Sustained disagreement from users caused models to reduce their accuracy by 30 to 35 percent. Most models changed their answers after the first round of pushback.

Language performance was also tested across 12 non-English languages using five models. While top models saw accuracy drops of about two percentage points, others saw declines of up to 10 percent in languages like Swahili.