Search

AI Models Struggle with Climate Facts Under User Pressure, UNESCO-UCL Report Reveals

Antalya:AI models, which initially provide correct answers to climate-related questions, tend to abandon established scientific facts when faced with user push-back. This finding is highlighted in a new report by UNESCO and University College London (UCL), presented at COP31.

According to Emirates News Agency, the report titled 'Can We Trust AI on Climate? Why it seeks to please the user, and fails on complex reasoning and multilingualism,' warns policymakers and researchers about the limitations of relying on AI models for developing climate policies. The report explains that AI models often deviate from scientific facts on climate change when challenged and struggle with processing new information, particularly in less widely spoken languages.

Researchers from UCL tested 15 large language models (LLMs) and found that while these models accurately answered neutral climate science questions, they tended to agree with inaccurate user assertions. The models pivoted from facts in 80% of cases when users mildly disagreed, leading to significant errors such as denying CO2's role as a primary driver of climate change.

Leona Verdadero, UNESCO's lead author, expressed concern over the design flaw in AI models that incentivizes agreement with users over factual accuracy. Dr. Maria Perez-Ortiz from UCL highlighted the risk of overestimating AI's capabilities in explaining scientific knowledge, emphasizing the need for caution when using AI to determine climate action.

The report also points out the disparities in AI model performance across languages. While models showed high accuracy in English, Spanish, French, and Portuguese, their accuracy significantly decreased in languages like Hindi, Bengali, Urdu, and Swahili. This is critical as regions speaking these languages are often more affected by climate change.

Popular Post

Pages