News Release

AI systems are easy to pressure into giving false climate information

Artificial intelligence models are easy to pressure into giving out false climate information despite being initially reliable, finds a new UNESCO-UCL report to be launched at an international conference

Reports and Proceedings

University College London

UCL Press Release
Under embargo until Monday 16 November 2026, 13:30 UK time / 16:30 Antalya time / 08:30 US Eastern time

Artificial intelligence models are easy to pressure into giving out false climate information despite being initially reliable, finds a new UNESCO-UCL report to be launched at an international conference.

The study, led by UCL researchers and published by UNESCO, examined the reliability of climate change information provided by 33 leading AI large language models, and tested five of them across 13 widely spoken languages, from English to Urdu. It will be unveiled at COP31 – this year’s meeting of world leaders, in Türkiye, at the annual United Nations’ Conference of the Parties climate change summit.

The researchers found that while these AI models could explain established climate facts reasonably accurately, the models were less reliable in languages with less digital text online, as well as when the AI was asked to make judgements, or when pressed by users to change their answers.

When asked to provide basic climate change facts, the best-performing AI systems responded with accurate information between 88% and 96% of the time, depending on the test. However, when asked to evaluate whether specific evidence supports or refutes a scientific claim, even the best models were only able to provide accurate answers between 71% and 77% of the time.

This accuracy dropped sharply when users expressed disagreement with climate facts and findings.

Across all of the 15 AI models tested in this part of the study, the accuracy of their climate information dropped by between 31 and 35 percentage points when users repeatedly disputed a factually correct answer, with about 80% of AI concessions occurring at the first sign of disagreement.

Every model tested gave up providing correct answers and adopted the user’s expressed incorrect biases after being pushed several times, an issue known as AI sycophancy.

Lead author Professor Ivana Drobnjak (UCL Computer Science and a member of the UNESCO Chair in AI at UCL) said: “While AI can help to make basic facts about climate change easier to understand in some circumstances, our research shows that there are clear limits to its abilities.

“It is not a dependable substitute for scientists and qualified experts. Its accessibility and conversational tone can give it an air of authority, but it cannot yet do the harder reasoning the climate problem demands, or hold on to a correct answer when a user pushes back.

“If AI is to help with climate change, it needs expert users who know how to question and check it, and AI systems built and tested specifically for the job.” 

The research team uncovered other issues with the reliability of climate information from AI models.

In the five models tested across languages, responses were more than 88% accurate in English, Spanish, French and Portuguese, but accuracy declined significantly for languages with less digital text online, falling to 83.6% in Bengali and Hindi, 81.4% in Urdu, and 78.3% in Swahili.

This resulted in AI models delivering often contradictory answers to the same scientific question when asked in different languages.

This also means that many of the communities facing the most severe climate impacts speak the languages that AI systems are least dependable in, and the researchers say that AI reliability may be even lower in languages with even less available digital text to draw on.

AI models also showed significant issues with evaluating the effectiveness of climate policies.

The team found that AI models could correctly identify how one specific policy on its own would affect global temperature by 2100 (e.g. whether a tax on coal would raise or lower it) up to 94.5% of the time.

However, when asked to combine two policies together and evaluate their outcomes (e.g. whether a coal tax and a natural gas tax combined would yield more or less reduction than the sum of each taken alone), the AI models answered with only 33.8% accuracy on average, no better than random chance (33.3%).

To test this, the team compared the AI responses against predictions from the En-ROADS climate simulator, a tool developed by Climate Interactive and MIT Sloan – the business school of the Massachusetts Institute of Technology in America – that is widely used by policy makers and in climate education to explore the effects of climate policies and technologies.

Senior author Dr Maria Perez-Ortiz (UCL Computer Science) said: “AI systems are remarkably good at explaining scientific knowledge, but this can give us a false sense of their capabilities.

“Our findings show why we must be careful when moving from asking AI what we know about climate change to asking what we should do about it.”

Co-author Leona Verdadero, Programme Specialist from UNESCO’s Digital Policies and Digital Transformation Section, said: “Recalling facts is not the same as being able to provide reliable policy analysis.

“The tendency for AI to misleadingly adapt answers to suit a user’s point of view, rely on weak evidence or present speculation as analysis shows its limitations as a trustworthy broker of climate information to the public.

“As more people come to rely on it as a source of information, it’s vital we understand what its shortcomings are.”

The research was carried out by the UNESCO Chair in Artificial Intelligence at UCL, a team of researchers which focuses on promoting ethical AI development, policy and education globally. The report was commissioned to contribute to UNESCO’s Global Initiative for Information Integrity on Climate Change.

Notes to Editors

For more information or to speak to the researchers involved, please contact Michael Lucibella, UCL Media Relations. T: +44 (0)75 3941 0389, E: m.lucibella@ucl.ac.uk

For journalists attending COP31, an in-person press conference will be held on 16 November at 16:00 Antalya time (13:00 UK time / 08:00 US Eastern time) at Press Conference room 1, Area D2-D3, Press Conference room Mevlana. A live webcast of the conference will be available at https://unfccc.int/event/unesco-launch-of-the-ai-and-climate-knowledge-report

 

Ivana Drobnjak, Leona Verdadero, Zekun Wu, Byambaa Bayarmandakh, Wietske Holwerda, Sindi Gjonaj, Maria Perez-Ortiz, ‘Can We Trust AI on Climate? Why it seeks to please the user, and fails on complex reasoning and multilingualism’ will be published by UNESCO on Monday 16 November 2026 at 13:30 UK time / 16:30 Antalya time / 08:30 US Eastern time and is under a strict embargo until this time.

Additional material

Links

 

About University College London (UCL) 

UCL is a global top 10 university, set up in London 200 years ago to offer education for all. Today, we gather 60,000 staff and students, from over 150 countries, to create a unique city within a city – a research and innovation powerhouse that leads the world in subjects spanning the arts, sciences, technology and the humanities. We’ve nurtured 33 Nobel Prize winners, because here, brave ideas have the scale and the support they need to succeed. We are University College London. And here, it can happen.  

UCL turns 200 in 2026. Join us for a year of bicentennial events and celebration.  

www.ucl.ac.uk 

 


Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.