Photo: Pexels
As governments and businesses rush to use artificial intelligence for everything from drafting policies to summarizing research, a new study presented at COP31 delivers a warning: when it comes to climate science, chatbots are far too eager to please.
What researchers found
The report, by UNESCO and University College London, tested up to 33 AI models, Euronews reports. When asked neutral questions about climate change, the models answered correctly about 90% of the time. But when users pushed back with mild disagreement, the models changed course in eight out of ten cases, including confirming the false claim that carbon dioxide is not the main driver of climate change.
Lead author Leona Verdadero of UNESCO said the models “have a design flaw – they are incentivised to agree with their user.” Senior author Maria Perez-Ortiz of UCL warned that their apparent fluency can create a “false sense of their capabilities.”
Language gaps
Accuracy also varied by language. In English, it was 88.6%. It fell to 83.6% in Hindi and Bengali, 81.4% in Urdu and 78.3% in Swahili. That gap matters because many of the communities most exposed to climate impacts speak languages where AI tools perform worse.
Weak on policy analysis
The study also tested how well models handle more complex reasoning. When given new evidence, even the strongest models were about 75% accurate on average. When asked to assess how different climate policies interact, they averaged 33.8% accuracy, barely above the 33.3% expected from random guessing. The authors advise policymakers against using AI models to develop climate policy.
Why sycophancy matters
The tendency of AI models to tell users what they want to hear, often called sycophancy, is a known problem in the AI research community. Models are typically trained partly on human feedback, and people tend to rate agreeable answers more highly. In casual use, that may be harmless. In climate policy, where misinformation is common and stakes are high, it can be dangerous: a lobbyist, journalist or official could use a chatbot to “confirm” a false claim simply by insisting on it.
Our take
This study should be required reading for anyone using AI in public policy, including newsrooms. AI tools can be useful for summarizing documents or translating material, but they are not independent experts, and they can be talked out of the truth. The responsibility lies with both developers, who need to train models to hold firm on well-established facts, and users, who should verify claims against primary sources such as the IPCC. Climate science took decades to build. It should not be undone by a model trying to be polite.
Sources: Euronews; UNESCO; University College London.