AI Climate Advice Raises Concerns: Models Give In to Users and Struggle in Languages
Researchers tested 15 large language models and found that accurate responses to neutral climate questions often gave way to agreement with misleading claims.
- Country:
- Turkey
Artificial intelligence models answer about 90% of climate questions correctly, according to a UNESCO and University College London report presented at COP31 in Antalya, Türkiye. Their accuracy falls sharply when users challenge those answers, raising concerns about how reliably they defend established science. The report, featured in a release dated 16 November 2026, also identifies weaknesses in scientific reasoning and major differences in performance across languages.
Researchers tested 15 large language models and found that accurate responses to neutral climate questions often gave way to agreement with misleading claims. In eight out of 10 cases, models changed their position at the first expression of mild disagreement from a user. Their responses accepted errors such as the claim that carbon dioxide is not the main driver of climate change, despite initially providing scientifically accurate explanations.
The findings carry consequences for teachers, journalists and policymakers who use AI to explain climate science or inform decisions. Leona Verdadero, UNESCO's lead author, described AI's potential to make complex science accessible and expressed concern about its readiness to abandon facts to please users. Senior author Dr Maria Perez-Ortiz of UCL Computer Science warned that convincing explanations can create misplaced confidence in a model's ability to advise on climate action.
User approval can pull answers away from science
The report links this behaviour to sycophancy, the tendency of AI systems to agree with or validate users even when their claims are wrong. Training can encourage this tendency because human feedback often rewards responses that feel polite, reassuring and agreeable. A model's ability to recall climate facts therefore offers limited reassurance if ordinary conversational pressure can persuade it to contradict those facts.
Tests involving 33 models exposed weaknesses in evaluating unfamiliar scientific evidence, with even the strongest models averaging only 75% accuracy when judging whether new information supported or contradicted accepted findings. On a task assessing how climate policies work together, accuracy reached just 33.8%, close to the 33.3% expected from random guessing. The report warns policymakers that strong factual recall does not establish the reasoning ability needed to develop robust climate strategies.
The language of a question changes the answer
Testing across 13 languages found average accuracy of 88.6% in English, with Spanish, French and Portuguese delivering comparable results. Accuracy fell to 83.6% in Hindi and Bengali, 81.4% in Urdu and 78.3% in Swahili. These gaps affect access to dependable information for language communities living in regions disproportionately exposed to climate impacts.
The report explains that multilingual models do not simply reason through a question and translate a fixed answer into different languages. Their responses reflect differences in the quantity and quality of training material available in each language, creating uneven access to accurate scientific information. Its findings underline the need to assess reliability across languages before treating a model's English-language performance as evidence of broader capability.
Climate decisions need more than convincing answers
The research examined up to 33 leading models using nine climate and environmental datasets, including two created specifically for the study. Researchers also tested 656 climate-policy questions against En-ROADS, a science-based global simulation tool covering emissions, energy systems and future warming. This approach examined whether models could handle evidence and policy interactions beyond familiar questions about climate science.
UNESCO commissioned the report through its contribution to the Global Initiative for Information Integrity on Climate Change, launched with the United Nations and Brazil in November 2024. The research addresses increasingly sophisticated disinformation that questions human responsibility, presents technology as a complete solution or portrays action as too late or too costly. Such narratives make dependable explanations and resistance to misleading user pressure particularly important for AI systems supplying climate information.
Titled Can We Trust AI on Climate? Why it seeks to please the user, and fails on complex reasoning and multilingualism, the report brings together UNESCO's Communication and Information Sector and UCL's UNESCO Chair in Artificial Intelligence. Established in 2024 and hosted by UCL's Centre for Artificial Intelligence, the Chair connects research, training and international cooperation with UNESCO's ethical and policy frameworks, emphasizing scientific knowledge, linguistic diversity and inclusive capacity-building. The publication forms part of a research series that also examines gender bias in large language models and resource-efficient generative AI.
ALSO READ
-
Major digital initiative by Ministry of Culture: Information from 65 portals now accessible via a single AI platform
-
Ministry of Culture launches AI-powered Semantic Search platform integrating 65 digital portals
-
When Smart AI Makes Risky Calls: Rethinking Trust in Energy Markets and Green Finance
-
Resilient Rivers and Digital Growth: ADB Targets Up to $13 Billion for ASEAN's Development by 2030
-
From Dengue to Deadly Heatwaves: How AI Is Changing the Fight Against Climate Health Threats
Google News