A study by UNESCO warns that artificial intelligence systems risk distorting climate science when used as public information sources. Testing 33 models with 656 questions showed that while basic science queries were answered correctly up to 96% of the time, performance on complex policy assessments dropped sharply to 33.8%. According to The Hindu BusinessLine, at higher complexity levels, the models performed no better than random chance. The report indicates that AI has become a new intermediary between science and the public, raising concerns about whether information arrives intact.
The research also highlighted a tendency for models to abandon established facts to align with user opinions. In tests with 15 models, researchers challenged correct answers with incorrect assertions, causing the systems to change their positions in eight out of ten cases. Leona Verdadero, a Programme Specialist at UNESCO, described this behavior as a serious threat to information integrity, noting that models prioritize being agreeable over maintaining factual accuracy. The study advises treating AI as a starting point rather than a substitute for expert knowledge.
Language barriers further reduce reliability. The study found that AI models were less accurate in Hindi, Bengali, Urdu, and Swahili compared to English, French, and Spanish. Accuracy fell by five to seven percentage points in the first group and by up to 10 percentage points for Swahili. This is critical because many speakers of these languages live in South Asia and East Africa, regions highly vulnerable to climate change. UNESCO emphasized that AI has become a new intermediary between science and the public, questioning whether information remains intact when filtered through these tools.
In response, UNESCO urged developers to invest in local-language datasets and improve scientific information availability in under-represented languages. The organization also called on governments to use public procurement to demand greater transparency and rigorous testing of AI systems. Guilherme Canela, Director of UNESCO’s Division for Digital Inclusion and Policies, stated that while AI can help millions understand climate change in their own language, it demands independent testing, transparency, and accountability. These recommendations were presented at a press conference during the COP31 climate negotiations on Monday.
