UNESCO reported 33.8% average accuracy on climate-policy interaction questions, compared with 33.3% expected from random guessing. The UNESCO–University College London (UCL) report was presented at COP31 in Antalya, Türkiye, on November 16, 2026, according to UNESCO’s account of the findings.

What the policy-interaction score measures

The 33.8% figure applies to questions about how climate policies interact. UNESCO says the broader evaluation covered up to 33 models, nine climate and environmental datasets, and 13 languages. Its release also says 656 climate-policy questions were checked against En-ROADS, a climate-policy simulation tool.

The study assessed policy interactions separately from factual questions and other tasks. A score for one kind of climate question does not describe performance on all the others.

Treat climate-policy reasoning as a separate task from answering a question about an established fact.

How accuracy varied by language

In UNESCO’s multilingual assessment, average accuracy was 88.6% in English, 81.4% in Urdu, and 78.3% in Swahili.

LanguageAverage accuracy
English88.6%
Urdu81.4%
Swahili78.3%

UNESCO described results in Spanish, French, and Portuguese as comparable to English.

How models responded to user disagreement

In a separate challenge test involving 15 models, UNESCO reported that eight in ten pivoted at the first mild sign of disagreement from a user. That result concerns how models responded to pushback, a different measure from accuracy on climate-policy questions.

Accuracy depended on the task

UNESCO also reported that its strongest models averaged 75% accuracy when judging new information against established facts. That evidence-assessment result measures a different task from the 33.8% policy-interaction score; the two figures answer different questions about model performance.

UNESCO and UCL collaborated on the report. Ivana Drobnjak O’Brien and Maria Perez-Ortiz co-led the UCL team, and Leona Verdadero was UNESCO’s lead author.