Technology
Ars Technica

"You’re right"—LLMs will get climate info wrong if you ask

Source Entity

Scott K. Johnson

November 21, 2026
"You’re right"—LLMs will get climate info wrong if you ask

A new study by UCL and UNESCO reveals that large language models frequently provide inaccurate information regarding climate science. The research highlights the risks of relying on ephemeral chat sessions for scientific data.

The Reliability Crisis in AI-Generated Climate Science

As the global discourse on climate change shifts from traditional search engines to conversational artificial intelligence, the integrity of information provided to the public has become a critical concern. Recent research conducted by University College London (UCL) in partnership with UNESCO has shed light on a troubling phenomenon: Large Language Models (LLMs) are systematically prone to providing inaccurate or misleading data when prompted with climate-related inquiries. While the quality of responses varies across different models, researchers have identified a consistent pattern of failure that undermines the public's ability to access reliable scientific consensus.

The Shift from Search to Synthesis

Historically, digital information consumption was dominated by search engines, which acted as gateways to static web pages. Researchers previously scrutinized these platforms by auditing the top results served to users. However, the current landscape has evolved toward LLMs, which synthesize information within private, ephemeral chat sessions. This shift creates a "black box" effect, where the information presented to an individual is unique to their interaction, making it difficult for researchers to audit the data diet of the general public. The UCL and UNESCO study represents a necessary evolution in media literacy research, attempting to quantify the accuracy of these automated systems.

Patterns of Failure in LLM Architectures

Despite advancements in natural language processing, LLMs operate based on probabilistic patterns rather than a deep, verified understanding of climate physics. The study suggests that while some models perform better than others, they all struggle with the nuances of climate science. This failure is likely tied to the models' training data, which may contain a mix of peer-reviewed research and climate misinformation. When an LLM generates a response, it may inadvertently prioritize linguistic coherence over scientific accuracy, leading to "hallucinations" or the presentation of outdated information that contradicts the current scientific consensus.

Implications for Public Climate Literacy

The reliance on AI for climate information has profound implications for public policy and individual behavior. If citizens receive inaccurate information from an LLM, their understanding of climate risks—such as extreme weather events, carbon sequestration, or renewable energy transitions—may be skewed. Because these interactions occur in private, there is no public accountability or correction mechanism, unlike traditional journalism or public social media discourse. This lack of transparency makes it increasingly difficult for educators and scientists to combat climate disinformation.

Future Trends and the Need for Verification

As LLMs become further integrated into our daily workflows, the need for robust verification protocols will become paramount. Future developments in AI must focus on grounding models in verified, real-time scientific databases rather than relying solely on the vast, unfiltered corpus of the internet. Without these guardrails, the risk of AI acting as a vector for climate misinformation will only grow. The UCL and UNESCO study serves as a vital wake-up call for developers to prioritize epistemic accuracy alongside conversational fluency in their quest to build more reliable and responsible AI systems.

Verification Required?

Read the full report from the primary source

Go to Ars Technica