When classifying climate misinformation, general-purpose large language models lag behind models trained on expert-curated climate data.