{
  "id": 20786012,
  "title": "LLMs will provide accurate climate info, except when they won’t",
  "url": "https://urgent.news/2026/11/19/llms-will-provide-accurate-climate-info-except-when-they-wont",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-19T18:44:39.000Z",
  "source": {
    "name": "Ars Technica",
    "slug": "ars-technica",
    "url": "https://arstechnica.com/science/2026/11/llms-will-provide-accurate-climate-info-except-when-they-wont/"
  },
  "original_language": "en",
  "account": "Many research studies have examined how individuals engage with information about climate science. Some investigations focused on search engine results, while others explored social media content or news coverage from newspapers and television. As large language models (LLMs) increasingly form part of the public information landscape, it is worth examining how these AI systems process climate change information. However, there is no straightforward method for researchers to observe the content individuals have encountered. This information is delivered temporarily to individual users during their chat sessions. A study led by researchers from University College London and UNESCO aimed to address this challenge by testing LLMs with various prompts and assessing their outputs. They began with simple factual questions, which could potentially be present in the models' training data, and also examined more complex interactions, including situations where users may disagree with or challenge an AI's answer. When conducting LLM research on CO2-related queries, critics often argue that the focus on a specific model is flawed, as other newer models may have been introduced after the experiment concluded. In this particular study, 33 models were utilized, encompassing various versions of major frontier models and open-weight models. The flagship frontier models examined were Claude Fable 5, Claude Opus 5, Gemini 3.1 Pro Preview, Gemini 3.6 Flash, and GPT-5.6 Sol. A comprehensive list of all the models tested can be found in the study's report.",
  "summary": "Some models fare better, but all fail in similar ways.",
  "key_points": [
    "Researchers from University College London and UNESCO tested 33 LLMs with various prompts.",
    "They examined both simple factual questions and complex interactions challenging AI answers."
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}