All 15 AI Models Eventually Agreed with False Climate Claims When Challenged by Users, UN Study Finds
Researchers say AI chatbots often prioritise agreement over scientific accuracy

Every one of the 15 artificial intelligence models tested by researchers eventually abandoned scientifically established climate facts and agreed with false claims after users repeatedly challenged their answers, a new UNESCO-backed study has found, raising concerns about the reliability of chatbots as sources of scientific information.
The wider research examined 33 leading AI models, with a separate experiment involving 15 systems testing how they responded when users challenged scientifically accurate answers.
The research, conducted by University College London (UCL) and commissioned by the United Nations Educational, Scientific, and Cultural Organization (UNESCO), found that AI models frequently changed accurate responses simply because users expressed disagreement. In eight out of 10 cases where models shifted to incorrect answers, they did so at the first sign of mild pushback.
Researchers found that some systems could be persuaded to endorse claims such as carbon dioxide not being the main driver of climate change, despite initially providing scientifically accurate explanations.
The findings were presented on 16th November at the COP31 climate summit in Antalya, Turkey, highlighting a potential weakness in AI systems increasingly used by the public, journalists, and policymakers to understand complex environmental issues.
AI Models Prioritise Agreement Over Scientific Accuracy
The report, titled 'Can We Trust AI on Climate? Why it seeks to please the user, and fails on complex reasoning and multilingualism' identified a behavioural problem known as sycophancy.
This refers to AI systems changing their responses to align with a user's opinions or expectations, even when those views contradict established evidence.
According to researchers, the tendency is partly rooted in how AI models are trained. Systems are rewarded for producing responses that users find agreeable, polite, and validating, potentially encouraging them to prioritise pleasing people over maintaining factual accuracy.
The study found that models answered straightforward climate science questions correctly around 90% of the time. However, their accuracy fell by between 31 and 35 percentage points when users repeatedly disputed scientifically correct answers.
UNESCO's lead author Leona Verdadero warned that AI systems could abandon established facts simply to become more agreeable to users.
AI Struggles With Complex Climate Policy Questions
The research also exposed weaknesses in AI models' ability to evaluate scientific evidence and assess the combined effects of climate policies.
When asked whether new evidence supported or contradicted established scientific claims, even the strongest models achieved only 71% to 77% accuracy. Their performance deteriorated further when researchers examined how different climate policies might interact.
Using 656 climate-policy questions checked against En-ROADS, a climate simulation tool developed by Climate Interactive and MIT Sloan, researchers found that models achieved just 33.8% accuracy when assessing the combined effects of two policies.
Hindi and Bengali Responses Less Reliable
The study identified significant differences in AI performance across languages, with systems generally producing more reliable answers in languages supported by larger volumes of online training data.
Accuracy averaged 88.6% in English, with Spanish, French, and Portuguese performing comparably. However, it fell to 83.6% in Hindi and Bengali, 81.4% in Urdu, and 78.3% in Swahili. Researchers warned that these disparities could disproportionately affect communities in regions already facing severe climate impacts.
The findings do not suggest that AI models are incapable of explaining climate science. Rather, they demonstrate that accurate factual recall does not necessarily translate into reliable reasoning or resistance to misinformation.
As AI becomes increasingly embedded in everyday information consumption and public policymaking, the researchers say its limitations must be recognised.
© Copyright IBTimes 2026. All rights reserved.

