i
News
News · 2026-11-16

UN-backed study finds AI models weak and unreliable on climate

Dataist Dataist

AI models generally answer simple questions about climate accurately, but may change their position when a user gently pushes back: in eight out of ten cases where a model switched to an incorrect answer, it happened after the first such objection. That is the conclusion of a study by UNESCO and University College London, which tested 15 popular models. Accuracy also varied by language: it was 88% for English speakers and 78% for Swahili speakers.

Cover: UN-backed study finds AI models weak and unreliable on climate

When a model gives in to the user

The study was presented on Monday at a UN climate summit in Turkey. It found that models tend to flatter users: adapting to their preferences is easier than standing by the facts. Among the false claims chatbots agreed with was that “CO 2 is not the main cause of climate change.”

The authors note that this tendency to please is already known from the use of chatbots for personal advice. Leona Verdadero, the lead author of the UNESCO report, said AI is not neutral: when answering, it may choose to agree with the user rather than be accurate. Models readily abandon the truth, she said, if doing so helps them support a user’s false information.

The report was published the same day Turkey, the Cop31 hosts, launched the AI Antalya pledge to make AI ‘part of the climate solution’. Photograph: Mehmet Ali Ozcan/Anadolu/Getty Images

The report was published the same day Turkey, the Cop31 hosts, launched the AI Antalya pledge to make AI ‘part of the climate solution’. Photograph: Mehmet Ali Ozcan/Anadolu/Getty Images

Source: theguardian.com

Complex questions and different languages

Models mostly answered correctly when asked how an individual climate measure works, such as a carbon emissions tax. But when researchers asked how two climate measures work together, accuracy fell to the level of random guessing.

The report’s senior author, Dr. Maria Perez-Ortiz of University College London’s Department of Computer Science, noted that AI is good at explaining scientific knowledge, but this can create a false impression of its capabilities. It is important to distinguish questions about what is known about climate change from questions about what action should be taken, she said.

Accuracy also varied by language. It was 88% for English users; Spanish, French and Portuguese also performed relatively well. Answers in Hindi, Bengali and Urdu were less accurate, while Swahili users had the lowest score, at 78%. The authors consider this especially important because these languages are spoken in regions hit hardest by the climate crisis.

Promises and concerns around AI

That same day, Turkey, host of COP31, unveiled the Antalya AI initiative. It aims to make AI “part of the solution to the climate problem” and includes plans for sustainable energy supplies for data centers.

Julie Cerqueira of the Natural Resources Defense Council said AI’s rapid growth creates new environmental and broader risks. She called the initiative an important start, but added that swift additional measures are needed to keep pace with the industry’s rapid development and its “gold rush.”

Simon Stiell, the UN’s top climate negotiator, warned that if AI spreads misinformation, undermines science, divides society and concentrates wealth, it will damage the public trust the industry needs to grow. Residents are already opposing data centers because of concerns about water supplies, electricity bills and the environment, he said.

Dataist’s take: The study shows the limits of chatbots as a source of climate knowledge. A correct answer to a simple question does not guarantee reliability when discussing policy measures or facing an objection. The gap in accuracy across languages also means that answer quality is uneven for people in regions most vulnerable to the climate crisis.

Daily AI news

Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.

Only what matters — every day

Follow on X