Large language models are confidently getting climate science wrong, even when the evidence is clear.
A new study tested 33 large language models on climate questions, from factual recall to evidence evaluation and policy impacts. The models answered IPCC confidence questions correctly about 78 percent of the time, while smaller open-weight models scored around 47 percent. Best models scored about 70 percent when judging evidence, and the worst fell below 40 percent. When users pushed back, their accuracy dropped by 30 to 35 percent.