UNESCO Report Finds AI Chatbots Often Cave to Pushback on Climate Questions

UNESCO Report Finds AI Chatbots Often Cave to Pushback on Climate Questions

A new UNESCO report finds that artificial intelligence chatbots generally get basic climate facts right, but frequently abandon correct answers when users repeatedly push back during a conversation.

The report, titled “Can We Trust A.I. On Climate Information?”, tested 15 different AI models. It found the chatbots performed well on single, straightforward questions, such as whether a carbon tax would reduce projected warming, reaching accuracy as high as 94.5%. But when two policies were combined in a single question, accuracy dropped sharply to 33.8%.

“Since meaningful climate action relies on overlapping policies working together, A.I. outputs cannot replace validated simulation models or expert analysis,” the report’s authors wrote.

The report also found that the quality of climate information dropped in “lower-resource languages” such as Hindi, Urdu, and Swahili — languages spoken in regions the Intergovernmental Panel on Climate Change (IPCC) has identified as among the most vulnerable to climate change.

Chatbots and ‘Sycophancy’

A central finding involves what the report calls “sycophancy”: the tendency of chatbots to change their answers in response to persistent disagreement from users, even when their original answer was correct.

Across the 15 models tested, UNESCO found a “mean drop of 31 to 35 points” in accuracy when users repeatedly challenged the chatbot’s answers.

“Larger and stronger systems generally outperform smaller ones, and current flagship models outperform earlier or lighter versions from the same developer,” the report said. However, it noted that even the most advanced models performed about as poorly as smaller ones when it came to “judging evidence” rather than simply reporting data.

“Models can be argued out of correct climate answers – mainly by what the user says, not by who the user claims to be,” the report said.

Notably, the report found that when a user claimed to be a climate scientist, all but two of the AI models tested were more willing to abandon a correct answer and accept the user’s incorrect claim instead.

What UNESCO Recommends

UNESCO recommended that “A.I. systems used in high-stakes climate-information contexts should be subject to independent evaluation, monitoring and accountability, covering factual accuracy, evidence verification, multilingual performance, robustness to manipulation, uncertainty communication, and policy reasoning.”

The report also called for greater “source transparency” from AI developers, so that the evidence behind a chatbot’s conclusions is presented in a format users can understand and verify themselves.

“While AI can help to make basic facts about climate change easier to understand in some circumstances, our research shows that there are clear limits to its abilities,” said Prof. Ivana Drobnjak of UCL Computer Science, a lead author of the report, on Monday.

“It is not a dependable substitute for scientists and qualified experts. Its accessibility and conversational tone can give it an air of authority, but it cannot yet do the harder reasoning the climate problem demands, or hold on to a correct answer when a user pushes back,” Drobnjak said.

Senior author Dr. Maria Perez-Ortiz added: “Our findings show why we must be careful when moving from asking A.I. what we know about climate change to asking what we should do about it.”

The findings echo broader concerns about AI reliability across other fields, including well-documented cases of chatbots producing fabricated legal citations in court filings. Experts have generally advised users to independently verify facts and claims generated by AI systems rather than accept them at face value.

Author picture

Leave a Replay

Recent Posts

Sign up for Joe's Newsletter, The Daily Informant