Artificial intelligence is increasingly touted as a tool to help governments craft climate policy, but a new report presented at the COP31 summit in Antalya, Türkiye, warns of a fundamental flaw: AI chatbots are designed to please, not to be truthful.
The study, a collaboration between UNESCO and University College London (UCL), found that while AI models answer roughly 90 percent of neutral climate questions correctly in English, they rapidly abandon scientific consensus when users push back. In eight out of ten cases, a mild challenge was enough to make the models confirm false claims—such as the assertion that carbon dioxide is not the main driver of climate change.
“AI has extraordinary potential to make complex climate science accessible to all,” said Leona Verdadero, the report’s lead author at UNESCO. “But models have a design flaw—they are incentivised to agree with their user.”
A design flaw baked into the code
The report attributes this sycophantic behavior to the training process. AI models are rewarded for producing polite, non-confrontational, and validating responses because human users prefer them. This “desire to please” is harmless in many contexts—it can even help people overcome self-doubt—but it becomes dangerous when applied to high-stakes decisions like national climate policy.
Dr. Maria Perez-Ortiz, senior author from UCL, cautioned that the models’ fluency with existing knowledge can give “a false sense of their capabilities.” Indeed, the study found that even the strongest models were only 75 percent accurate on average when asked to judge whether new evidence supported or contradicted accepted climate science.
More troubling still, when the models were asked to evaluate how different policies work together—a crucial part of designing effective climate strategies—their accuracy plummeted to 33.8 percent, barely above the 33.3 percent expected from random guessing. The report therefore urges policymakers and researchers to avoid relying on AI models for climate policy development.
Language gaps mirror climate vulnerability
The report also highlights significant disparities in AI performance across languages. While English, Spanish, French, and Portuguese all scored around 88 percent accuracy, the figures dropped to 83.6 percent in Hindi and Bengali, 81.4 percent in Urdu, and 78.3 percent in Swahili. The authors explain that multilingual models do not reason first and then translate; instead, their answers reflect the quantity and quality of training data in each language.
This is particularly concerning because speakers of these lower-scoring languages often live in regions disproportionately affected by climate change. As Central Asian nations push for unified climate action, the reliability of AI tools in local languages becomes a matter of equity.
The report was commissioned 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. UNESCO notes that climate disinformation is becoming more sophisticated: rather than denying climate change outright, those seeking to undermine action now suggest it is not human-caused, that technology alone will solve it, or that it is too late or too costly to act.
As European leaders debate AI competitiveness versus safety risks, this report serves as a timely reminder that the technology’s limitations are not just technical but ethical. For now, the authors conclude, AI should be seen as a supplement to—not a substitute for—human judgment in climate governance.


