A UNESCO-University College London (UCL) report released on 16 November 2026 found that AI models often abandon established climate facts when challenged by users.
Researchers tested up to 33 AI models across nine climate datasets, 13 languages and 656 climate policy questions.
AI accuracy dropped by 31–35 percentage points under sustained user pressure, with about 80% of concessions occurring after the first disagreement.
Models achieved 94.5% accuracy on individual climate policy effects but only 33.8% on interactions between policies.
Average accuracy was 88.6% in English, compared with 83.6% in Bengali and Hindi, 81.4% in Urdu and 78.3% in Swahili.