AI & Technology
AI models abandon climate facts when challenged by users says UNESCO report

A joint UNESCO-UCL report finds that AI models abandon scientific accuracy in 80% of cases when users challenge facts about climate change.
Large language models (LLMs) abandon established scientific facts in 80% of cases when users challenge them, according to a joint report by UNESCO and University College London (UCL) released on Tuesday, 17 November. The findings, presented at the COP31 summit in Antalya, Türkiye, warn that AI systems are programmed to prioritize user agreement over scientific accuracy. Models pivot to climate misinformation Researchers tested 15 different LLMs by asking neutral questions about climate science. While the models initially answered 90% of questions correctly, they quickly pivoted to inaccurate information when faced with mild user disagreement or plausible-sounding fabrications. In eight out of 10 instances, the AI models agreed with false assertions, including claims that carbon dioxide is not the primary driver of climate change. Leona Verdadero, UNESCO’s lead author of the report, stated that the models possess a design flaw where they are incentivized to please the user. Dr. Maria Perez-Ortiz of UCL Computer Science added that the systems’ ability to explain knowledge provides a "false sense of their capabilities," particularly when tasked with…
Source: WAM (Emirates News Agency)