Researchers from University College London and UNESCO tested 33 large language models on climate change questions and found all models fail in similar ways. The study evaluated flagship models including Claude Opus 5, Gemini 3.1 Pro Preview, and GPT-5.6 Sol alongside smaller open-weight models.

On basic factual recall, frontier models and large open-weight models scored around 90 percent on multiple-choice and reading comprehension questions. Small open-weight models scored in the 70s. Questions about scientific uncertainty proved hardest, with flagship models scoring around 78 percent and the smallest models at 47 percent.

The models performed worse when asked to evaluate whether evidence supported specific claims, with top performers reaching 70 percent accuracy while worst performers scored below 40 percent. On questions about how climate policies would impact global temperatures, single-policy scenarios were answered correctly 72 to 89 percent of the time. But when asked how pairs of policies would combine, models performed no better than random guessing.

Models proved vulnerable to user pushback. When researchers presented sustained disagreement and counter-arguments over three rounds, all 15 models tested reduced their accuracy by 30 to 35 percent on false climate claims. Most models changed their answers after just one prompt of disagreement.

Testing in 12 non-English languages showed accuracy drops. While top models dropped only about two percentage points, lower-performing models fell seven percent (Urdu) or 10 percent (Swahili). GPT-5.6 Sol dropped 3 percent in Swahili while DeepSeek-V3 dropped 16 percent.

The researchers noted frontier models can answer textbook climate questions and produce explanations scientists judge comparable to PhD-level work. However, they wrote, "this knowledge is uneven: models are less reliable at judging evidence and scientific uncertainty; that knowledge is not equally accessible across languages; it is not always robust to user pressure; and models fail dramatically when asked to reason about how climate policies interact."