Over the past few years, generative AI has been heavily deployed in numerous fields of science, though with plenty of caution in tow. A new United Nations-backed report suggests that it’s pretty easy to convince an AI model to give out false climate information. And it’s not just about AI models giving out hallucinated knowledge. The more worrying part is that even when they give out correct information, a little bit of goading will have them doling out wrong climate-related information.
Now, this isn’t an isolated case. In 2025, an Earth Day article highlighted how AI chatbots can produce inaccurate or made-up answers on climate topics, citing a study in which Google’s Bard, now Gemini, produced falsehoods on every climate-related narrative it was tested with. A separate investigation by Global Witness found that chatbots, including ChatGPT, Meta AI, and Grok, sometimes shared climate disinformation tropes and amplified climate denial influencers.
The Stockholm Resilience Centre has also warned that generative AI, combined with social media algorithms and bots, could create a “perfect storm” of climate misinformation. Furthermore, research presented at the AAAI Conference on Artificial Intelligence in Philadelphia in 2025 found that open models such as Meta’s Llama did a shoddy job of identifying misleading climate claims, scoring as low as 0.28 on a 1.00 scale, while GPT-4 scored 0.74.
Sycophancy continues to be the big problem
The latest flub in the series comes courtesy of UNESCO and University College London. The researchers assessed 33 AI models, and tested five of them across 13 languages, each with millions of speakers across the globe. They found that these AI models are generally pretty accurate at spitting out factual information about climate. But when the same conversation was steered into a language with a smaller corpus of text on the internet, accuracy dipped. Likewise, when users pressed these AI models, they flipped on the accurate information and agreed with the wrong information.
On basic facts, the best models were accurate between 88% and 96% of the time. But when asked to judge whether evidence supports or refutes a claim, even the best models’ accuracy dipped to between 71% and 77%. When users repeatedly disputed a correct answer, the problem worsened further, with accuracy falling by 31 to 35 percentage points. About 80% of the time a model gave in, it folded at the very first sign of disagreement. And every single model in the pressure test eventually adopted the user’s incorrect view after repeated pushback.
In the realm of AI, this behavior is known as sycophancy, where an AI model tells users what they want to hear, even if it means agreeing with wrong beliefs. Complex questions were another weak spot. When asked how a single climate policy would affect global temperatures in 2100, models got it right up to 94.5% of the time. But when policies were combined, average accuracy plummeted to 33.8%, roughly what you’d expect from random guessing.
Language is another crucial pitfall

Language proved to be another weak point of these AI models when it comes to discussing climate-related topics. When questions were posed in English, Spanish, French, and Portuguese, accuracy was above 88%. But in Bengali and Hindi, accuracy dropped to 83.6%, falling to 81.4% in Urdu and 78.3% in Swahili. Now, these are no rare languages. Bengali, Hindi, and Urdu are collectively spoken by hundreds of millions of people across the globe.
The problem is that digital information in these languages isn’t as widely available as it is in English and other European languages. This isn’t the first research of its kind where AI models have stumbled as soon as they shifted to a non-English language. Even when responding to the same climate-related question in different languages, the AI sometimes served contradictory answers.
The report’s lead author, Prof. Ivana Drobnjak of UCL Computer Science, notes that AI “is not a dependable substitute for scientists and qualified experts.” UNESCO’s Leona Verdadero, a co-author, adds that “recalling facts is not the same as being able to provide reliable policy analysis,” a warning worth keeping in mind for anyone consulting AI about climate strategies and action plans.
