AI models abandon climate facts when challenged, UN-backed study finds
Popular AI models readily shifted from accurate climate information to misinformation when users challenged them, according to a study by Unesco and University College London. Tests of 15 models found a strong tendency to prioritise agreement with users over factual accuracy. Among instances in which models switched to a wrong answer, eight times out of 10 they did so at the first sign of mild disagreement.
The systems were mostly accurate when explaining how a single climate policy, such as a carbon tax, would work. But accuracy fell to no better than random guessing when they were asked how two climate policies could work together. Researchers warned that fluent explanations of established science could create misplaced confidence in the systems’ ability to guide climate decisions.
Performance also varied by language. Accuracy reached 88% in English but fell to 78% in Swahili, with weaker results also recorded in Hindi, Bengali and Urdu. Researchers highlighted the implications for communities particularly exposed to climate impacts. The findings were released at the UN climate summit in Turkey as the host launched the AI Antalya pledge to make AI part of climate solutions. UN climate chief Simon Stiell warned that disinformation and environmental concerns surrounding datacentres could undermine public trust.