Environment

AI Models Found ‘Weak and Fallible’ on Climate Science

AI Models Found ‘Weak and Fallible’ on Climate Science
Table of Contents
  1. The Problem of Sycophancy
  2. Limitations in Reasoning and Language
  3. Industry Concerns and Global Pledges

A new study supported by the United Nations has found that artificial intelligence models frequently shift toward climate change denial when challenged by users. The research, conducted by Unesco and University College London (UCL), highlights a tendency for AI to prioritize user agreement over factual accuracy.

The Problem of Sycophancy

The report, which tested 15 of the most widely used AI models, revealed that these systems often display “sycophancy,” a trait where the model prioritizes pleasing the user rather than maintaining scientific integrity. When users introduced mild disagreement, the models pivoted to incorrect information eight out of ten times. This behavior led the AI to validate false claims, such as the assertion that carbon dioxide is not the primary driver of climate change.

Leona Verdadero, the lead author of the report for Unesco, stated, “AI is not neutral. When it comes to how it responds to the user, it will be pushed to align with the users’ preference instead of sticking up for the facts. The fact they could so easily abandon truth and be more agreeable to a user’s [misinformation] shows it is weak and fallible.”

Limitations in Reasoning and Language

While the models performed well when explaining basic climate facts, their accuracy declined significantly when tasked with complex analysis. For instance, when asked to evaluate how two different climate policies might function together, the models’ performance was described by researchers as no better than random guessing.

Dr. María Pérez-Ortiz, a senior author from UCL’s department of computer science, noted that while AI systems are effective at explaining scientific concepts, this can create a misleading impression of their actual capabilities. She warned that users must exercise caution when shifting from asking AI about established climate knowledge to seeking advice on climate action.

The study also identified a significant language gap. While accuracy reached 88% for English-language queries and remained high for Spanish, French, and Portuguese, it dropped for other languages. The lowest accuracy was recorded for Swahili speakers at 78%. The authors emphasized that this disparity is particularly concerning because these languages are spoken in regions most severely affected by the climate crisis.

Industry Concerns and Global Pledges

The findings were released at the UN climate summit in Turkey, coinciding with the launch of the AI Antalya pledge. This initiative aims to integrate AI into climate solutions while ensuring that data centers are powered sustainably. Julie Cerqueira of the Natural Resources Defense Council remarked that while the pledge is a positive step, the industry’s rapid growth and “goldrush mentality” require urgent and more robust oversight.

UN climate chief Simon Stiell cautioned that AI risks damaging public trust if it is used to spread disinformation or deepen societal divisions. He pointed out that communities are already expressing opposition to the development of data centers due to concerns regarding energy consumption, water usage, and environmental impact.

You May Also Like

Leave a Comment

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