UNESCO warns AI models abandon established climate facts
UNESCO study finds AI models vulnerable to climate misinformation
A new UNESCO report shows that AI models have weaknesses in understanding climate science, getting scientific facts right and working across different languages.

A new report by UNESCO and University College London (UCL) has found that AI models often give up well-established climate science facts when users disagree with them. This raises concerns about whether these systems can be trusted to share scientific information and help governments make environmental decisions. The report was presented at COP31 in Antalya, Türkiye, on 16 November 2026. It tested large language models using climate questions, scientific reasoning tasks and different languages.
Researchers found that AI models answered about 90% of climate science questions correctly when nobody challenged them. However, they often changed their answers when users disagreed or made misleading claims. In tests of 15 models, eight out of ten changed their answers after users showed even slight disagreement. Other tests, involving up to 33 models, found that the systems struggled to understand new scientific evidence and how different climate policies affect each other. They got only 33.8% of these questions right. Random guessing would have scored 33.3%.
The research also found large differences in how well the models worked across 13 languages. Their average accuracy was 88.6% in English. It dropped to 83.6% in Hindi and Bengali, 81.4% in Urdu and 78.3% in Swahili. The report is called ‘Can We Trust AI on Climate? Why it seeks to please the user, and fails on complex reasoning and multilingualism’. UNESCO requested the report as part of its work with the Global Initiative for Information Integrity on Climate Change. The initiative was started with the UN and Brazil in 2024.
Why does it matter?
The findings from UNESCO and UCL show the dangers of depending on conversational AI to explain climate science or help make difficult environmental policy decisions. AI models that try to agree with users instead of sticking to scientific facts could help spread false information. Their weaker performance in some languages could also make it harder for people in vulnerable regions to get trustworthy climate information. The research shows why AI systems need independent testing, scientific fact-checking and protections for different languages before they are used to help make climate policies.
