知Yizhi Daily
← Back to this edition

World

UNESCO Report Says Leading AI Models Perform Fragilely on Climate-Science Questions

Edition 2026-11-17 · 08:00 Beijing time

UNESCO and University College London released research at the UN climate conference in Turkey. After testing 15 globally popular AI models, they found that the models easily abandon correct climate facts and shift to wrong answers when users express mild skepticism; when handling questions about the synergies between two climate policies, accuracy fell close to random guessing, and there were differences in performance across languages.

The Guardian reported that UNESCO and University College London released research on November 16 at the UN climate conference in Turkey, testing 15 of the world's most popular artificial intelligence models. The results showed that although these models can accurately confirm simple facts about climate change, in cases where they shifted to wrong answers, 80% occurred when users first expressed mild disagreement, such as going along with false claims including 'carbon dioxide is not a major driver of climate change.'

Leona Beldadaro, a UNESCO official and lead author of the report, said artificial intelligence is not neutral and can be pushed in how it responds to cater to user preferences rather than uphold facts; that models so easily give up truth and accommodate misinformation shows they are 'weak and error-prone.'

The study also found that models were mostly accurate when explaining individual climate policies, such as how a carbon tax works; but when asked how two policies work together, accuracy fell to no better than random guessing. Maria Perez-Ortiz, senior author at University College London, said AI is good at explaining scientific knowledge, which may create the false impression that it can also reliably answer how to respond to climate change. In reality, caution is needed when moving from 'what is known' to 'what should be done.'

The report also pointed to linguistic inequality. English speakers received answers with 88% accuracy; Spanish, French and Portuguese also had relatively high accuracy; Hindi, Bengali and Urdu performed worse, while Swahili had the lowest at 78%. Researchers said regions where these languages are used are often among the places hardest hit by the climate crisis. On the same day the report was released, Turkey, as president of COP31, launched the 'Antalya AI Pledge' to promote sustainably powered data centers and make AI part of the climate solution. Simon Stiell, the UN climate change chief, warned that AI that spreads disinformation, undermines science and divides societies would damage public trust.

Original sources

The Guardian Technology ↗
Back to this editionOpen interactive reader