AI Models Falter on Climate Science Questions
A study of 33 large language models found weaker accuracy on climate evidence, policy interactions and under user pushback.
Updated on Nov. 19, 2026 in Environmental

Live Poll
Do you trust AI chatbots to give you accurate information on science topics?
Researchers tested 33 large language models on 600 climate questions and found weaker performance on evidence, policy interactions and user pushback. Under sustained user disagreement, 15 tested models lost 30% to 35% accuracy.
Why it matters
As people increasingly turn to chatbots for information on climate change, the study suggests answers can be unreliable depending on model size, question type and the language used. Accuracy on policy interactions was no better than random guessing, raising questions about how such tools inform public understanding.
Frontier and large open-weight models generally scored around 90% on other climate question sets, but the strongest models reached only about 70% when judging whether evidence supported or refuted specific claims. The study does not state how these results would change with newer model releases.
The details
Researchers tested models on benchmarks covering IPCC confidence, evidence evaluation and policy effects on future warming. Scores also fell in languages other than English, including a seven-point decline in Urdu and a 10-point decline in Swahili, and about half of the questions received different answers depending on the language used.
Big Picture
The study measures how well language models reflect IPCC confidence levels, which are the consensus assessments underpinning most climate policy discussion. Its results suggest that reliance on these models for climate information may carry risks that current benchmarks only partly capture.
Readers who use chatbots for climate questions should verify key claims against primary sources, particularly on evidence evaluation and policy effects. Accuracy varied by language and by how strongly a user pushed back on an answer.
The takeaway
Language models can sound confident while missing key climate evidence, so checking answers against established scientific assessments remains essential.
Further reading
Explore more on Environmental research and how emerging technologies intersect with climate science.
Source note: This article includes information reported by TokenPost.
Live Poll
Do you trust AI chatbots to give you accurate information on science topics?



