Can we trust AI to guide climate policy? New report warns against glaring ‘design flaw’

Can we trust AI to guide climate policy? New report warns against glaring ‘design flaw’

Published on 16 November 2026 - 14:58 GMT+1 The desire to please is baked into AI chatbots, which are designed to keep users engaged in conversation for as long as possible. ADVERTISEMENT ADVERTISEMENT This often means predicting and modelling what you want to hear – even if it means bending the facts. While it might be harmless for certain tasks – and can even be helpful for things like overcoming self doubt – it could be far more damaging if its influence seeps into high-stakes realms like government climate policy. AI chatbots can explain climate science well, but they quickly abandon the facts when users push back, according to a new report by UNESCO and University College London (UCL) presented at the COP31 climate summit in Antalya, Türkiye. Researchers found that AI models answered around 90 per cent of neutral climate questions correctly. But when users challenged those answers, or asserted plausible-sounding fabrications, the models rapidly gave way. In eight out of 10 cases, they changed course at the first sign of mild disagreement, including by confirming false claims such as that CO2 is not the main driver of climate change. ‘Sycophantic by design’: A design flaw that could skew climate policy The report says sycophantic behaviour is built in during training. Models are rewarded for polite, non-confrontational and validating answers, because humans prefer them. “AI has extraordinary potential to make complex climate science accessible to all,” says Leona Verdadero, UNESCO’s lead author of the report. But “models have a design flaw – they are incentivised to agree with their user”, she says. Senior author Dr Maria Perez-Ortiz of UCL adds that AI models’ fluency with existing knowledge can give “a false sense of their capabilities”. Despite having impressive factual recall, the models were found to be an unreliable tool for building robust climate policies. When they were given a new piece of evidence and asked whether it supported or contradicted accepted climate science, even the strongest ones were only 75 per cent accurate on average. Their performance was even worse when they were asked to assess how well policies work together – an important part of designing effective climate strategies. The models scored just 33.8 per cent accuracy on average, only slightly above the 33.3 per cent expected from random guessing. The report therefore warns policymakers and researchers against using AI models to develop climate policy. AI accuracy falls among some languages Performance also varied widely by language. Models answered 88.6 per cent of questions correctly in English, and did comparably well in Spanish, French and Portuguese. Accuracy fell to 83.6 per cent in Hindi and Bengali, 81.4 per cent in Urdu and 78.3 per cent in Swahili. That’s because multilingual models don’t reason first and then translate, the report says. Their answers reflect the quantity and quality of the training data in each language. Speakers of the lower-scoring languages live in regions disproportionately affected by climate change, the authors note. Climate disinformation is growing more sophisticated 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 says climate disinformation is becoming more sophisticated. Rather than denying climate change is real, those seeking to undermine climate change action increasingly suggest it isn’t human-caused, that technology alone will solve it, or that it's too late or costly to take action. The researchers tested up to 33 leading models on nine datasets, across 13 languages and 656 climate-policy questions, which they checked against En-ROADS, a science-based simulation tool.

Original Source

Read the full article at Euronews →

KhanList aggregates and links to publicly available news content. We do not host full articles from third-party sources. Always verify important information with original sources.