AI models abandon climate facts when users push back, UNESCO and UCL study finds
AI models are mostly accurate on simple climate facts but quickly switch to wrong answers when users push back, according to a report from UNESCO and University College London launched at the COP31 climate summit in Turkey, The Guardian reports. The researchers tested 15 of the world's most popular AI models.
Eight times out of 10 when a model pivoted to a wrong answer, it did so "at the very first sign that the user mildly disagreed", for example confirming that "CO2 is not the main driver of climate change". The report found a strong tendency to sycophancy. Models were mostly accurate when asked how a single climate policy such as a carbon tax would work, but when asked how two policies would work together, accuracy fell to no better than random guessing.
"AI is not neutral," said Leona Verdadero, UNESCO's lead author of the report. "The fact they could so easily abandon truth and be more agreeable to a user's misinformation shows it is weak and fallible."
We hope you enjoyed this article
Consider subscribing to one of our newsletters like AI Policy Brief or Daily AI Brief.
Also, consider following us on social media:
More from AI Safety
Nov 16 OpenAI embeds six Apollo Research evaluators to test whether models try to subvert safety controls Nov 16 OpenAI releases CSAM detection classifier and adds ChatGPT warnings and temporary bans Nov 16 OpenAI considered buying Anthropic-backed Goodfire for more than $5B, report says Nov 16 Anthropic says it has met its safety threshold for AI that speeds up AI research Nov 16 US and Russia set to oppose talks on binding autonomous weapons rules at UN meetingAI Policy Brief
Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.
Whitepaper
Tensordyne Napier: What If One Rack Could Do the Work of Nine?
Tensordyne
This Tensordyne whitepaper presents Napier, an inference-focused AI processor and rack-scale system based on the company’s TDN Math logarithmic number system. It examines infrastructure requirements for large mixture-of-experts and agentic models, compares major inference architecture approaches, and details the TDN AIP processor, TDN72 pod, TDN Link fabric, and Napier Ultra configuration. The paper reports simulation-based performance, cost, and accuracy-validation results, including Tensordyne’s projected comparison of one Napier rack with a nine-rack Nvidia Rubin plus Groq deployment; the chip is reported as taped out and in fabrication.
Read moreYou may also like
Study Finds Errors in 40% of AI Chatbot Election Answers
MIT Trains Climate AI With Less Simulation Data
AI models give sound legal basics but miss local help in legal aid test of 23 models
Anthropic safety guide shows Claude Fable 5.1 scored 94% in app but 60% via API on self-harm tests
Ai2's XYBench tests whether AI assistants catch misguided questions
Daily AI Brief: the AI news that matters, in your inbox.