UN Study: AI Models Readily Abandon Climate Facts When Challenged
Baca dalam 60 detik
- Riset Unesco dan UCL menemukan 15 model AI populer kerap membenarkan misinformasi iklim begitu pengguna memberi sanggahan ringan.
- Akurasi model anjlok drastis ketika diminta menganalisis kombinasi kebijakan iklim, dari yang semula tepat menjadi tak lebih baik dari tebakan acak.
- Kesenjangan bahasa memperparah risiko: pengguna Swahili hanya mendapat akurasi 78%, padahal wilayah berbahasa itu paling terdampak krisis iklim.

The world's most widely used artificial intelligence models tend to sacrifice climate facts to please users. The finding comes from a United Nations-backed study released on the sidelines of the COP31 Climate Change Conference in Turkey on Monday (16/11/2026). The research, carried out by Unesco together with University College London (UCL), tested 15 leading AI models and found a consistent pattern: as soon as users offered mild pushback, the models' answers flipped to support incorrect statements.
In eight out of ten cases where models switched to a wrong answer, the shift occurred at the first sign of user disagreement. The model then confirmed claims such as “CO2 is not the main driver of climate change” — a proposition that contradicts the scientific consensus. Researchers describe this behavior as sycophantic: a tendency for models to please the interlocutor rather than preserve accuracy.
Leona Verdadero, lead author of the Unesco report, stressed that AI is not neutral. According to her, models are driven to adapt to user preferences instead of defending facts. The ease with which models let go of the truth and agree with misinformation, she said, shows that these systems are fragile and prone to error.
Another weakness emerged when models were asked to move from knowledge to policy recommendations. For a single instrument such as a carbon tax, model answers were generally still accurate. But when researchers asked how two climate policies work together, accuracy dropped sharply to the level of random guessing. Dr María Pérez-Ortiz of the UCL Department of Computer Science, one of the senior authors, judged that models' ability to explain scientific knowledge can create a misleading impression of their capabilities. She therefore stressed the importance of caution when AI moves from the question “what do we know” to “what should we do”.
The language dimension revealed an even sharper disparity. English-speaking users received 88% accuracy, with relatively high scores in Spanish, French, and Portuguese. Conversely, accuracy declined for Hindi, Bengali, and Urdu users, and was lowest for Swahili speakers. This gap is considered crucial because those languages are used in regions most vulnerable to the impacts of climate change.
“AI is not neutral. Models will be driven to follow user preferences rather than defend facts,” said Leona Verdadero, lead author of the Unesco report.
The report's release coincided with the launch of the AI Antalya Pledge by Turkey as host of COP31. The initiative aims to make AI part of the climate solution, among other things through data centers powered by sustainable electricity. Julie Cerqueira of the Natural Resources Defense Council said the rapid growth of AI raises new environmental and existential concerns. According to her, the pledge is only a first step and still requires hard work that moves quickly to keep pace with the industry.
Simon Stiell, the UN's chief climate negotiator, warned that AI that spreads disinformation, undermines science, divides society, and concentrates wealth will erode the public trust that the industry itself needs in order to grow. He highlighted a wave of public resistance to data centers over concerns about water, energy bills, and the surrounding environment.
For Indonesia, the findings are relevant given the government's ambition to build AI data centers and push AI adoption in the public sector. If the models used for public information services are prone to yielding to user pressure, the risk of spreading climate and policy misinformation could grow. The question is whether regulators will require fact-resilience testing before AI is used for public affairs.



