AI models tend to agree with users over climate facts: UNESCO

Researchers tested 15 large language models (LLMs) and found they generally provided accurate answers to neutral questions about climate science. However, when researchers challenged the answers or asserted plausible-sounding but false claims, the models quickly changed their positions, according to the report.

The report described this tendency to agree with users as “sycophancy” and linked it to training methods that reward polite, non-confrontational, and validating responses.

The study also found that while AI models performed well when recalling existing scientific knowledge, but their ability to analyze new scientific evidence remained limited. Performance also varied significantly across languages.

Maria Perez-Ortiz, a co-author of the report, said users should be particularly cautious when moving from questions about established climate knowledge to questions about what action should be taken.

People visit the 2025 World Internet Conference (2025 WIC) Light of Internet Expo in Wuzhen, east China’s Zhejiang Province, Nov. 6, 2025.(Xinhua/Huang Zongzhi)

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *