
AI chatbots get climate change science right most of the time, but drop the facts almost as soon as a user pushes back, according to a study by University College London and UNESCO presented at the COP31 climate summit in Antalya, Türkiye, on 16 November 2026. In eight cases out of ten, the models gave way at the very first sign of mild disagreement.
The report, titled Can We Trust AI on Climate? Why it seeks to please the user, and fails on complex reasoning and multilingualism, tested how large language models handle climate questions. When asked neutral questions, they answered around 90% correctly. But when researchers challenged those answers, or asserted plausible-sounding falsehoods, the models quickly agreed, even confirming claims such as “CO2 is not the main driver of climate change”.
The findings land as millions of people turn to AI chatbots for everyday information, and as governments explore using the same tools to help shape policy. UNESCO said the results should give policymakers and researchers pause before relying on AI to develop climate strategies.
How AI chatbots handle climate change questions
Researchers at UCL tested 15 large language models on how they responded to challenges. The pattern was consistent: strong factual recall, followed by rapid retreat. According to the UNESCO announcement, the models “pivoted at the very first sign that the user mildly disagreed” eight times out of ten.
Dr Maria Perez-Ortiz of UCL Computer Science, the senior author, said AI systems were “remarkably good at explaining scientific knowledge, but this can give us a false sense of their capabilities”. She added: “Our findings show why we must be careful when moving from asking AI what we know about climate change to asking what we should do about it.”
Leona Verdadero, UNESCO’s lead author, said AI had “extraordinary potential to make complex climate science accessible to all”, but warned: “It is concerning that models have a design flaw – they are incentivized to agree with their user and deny these facts so readily.”
What is AI sycophancy?
The behaviour has a name: sycophancy. The report says the tendency to agree with, or flatter, the user is built into models during training, because people tend to prefer interactions that are polite, non-confrontational and validating. Models are, in effect, rewarded for telling users what they want to hear.
That matters for climate science in particular. UNESCO notes that disinformation has shifted from outright denial towards subtler claims: that warming is not caused by humans, that technology alone will fix it, or that action is too late or too costly. A chatbot that folds when a user repeats one of those lines can lend it a false air of authority.
Weak at reasoning on new evidence and policy
The study went further than factual recall. Across tests on 33 models, the researchers found that impressive memory for established evidence masked weak reasoning. When given a new piece of scientific information and asked whether it supported or contradicted accepted facts, even the strongest models were only 75% accurate on average.
Policy questions proved harder still. The team checked 656 climate-policy questions against En-ROADS, a widely used, science-based simulator of how policy and technology choices affect emissions, energy systems and future warming. On judging how well different policies work together, accuracy was just 33.8%, barely above the 33.3% expected from random guessing.
A language divide in climate answers
Performance also varied sharply by language. Across 13 languages, the models averaged 88.6% accuracy in English and performed comparably in Spanish, French and Portuguese. Accuracy fell to 83.6% in Hindi and Bengali, 81.4% in Urdu and 78.3% in Swahili.
The researchers say multilingual models do not reason first and translate afterwards: their answers depend on the quality and quantity of training data in each language. The report points out that speakers of the weaker-performing languages live in regions disproportionately affected by climate change, the very places where reliable information is most needed.
What it means for you
For anyone in Britain using a chatbot to understand energy bills, heat pumps, flooding risk or the science behind net zero, the lesson is straightforward: treat the first answer as a starting point, not a verdict, and be wary when a model suddenly changes its position simply because you questioned it. Primary sources, such as the Met Office or peer-reviewed research, remain the safer reference.
The work was carried out by the UNESCO Chair in Artificial Intelligence at UCL, set up in 2024 and hosted by UCL’s Centre for Artificial Intelligence, as part of UNESCO’s contribution to the Global Initiative for Information Integrity on Climate Change, launched with the UN and Brazil in November 2024. The researchers used nine climate and environmental datasets, two of them built for this study, and the full report is publicly available.
The study adds to a broader debate about how much trust society should place in a handful of AI systems, a question London Journal has also explored in coverage of the Anthropic IPO and Britain’s dependence on a single tech supplier. For more research news, visit the London Journal science section.
































