UNESCO study finds AI models yield on climate facts
A UNESCO and University College London study found that all 15 artificial intelligence models tested in a user-pressure experiment eventually abandoned correct climate-science answers after repeated challenges, while models also scored close to random guessing on questions about combined climate policies.
The report, titled “Can We Trust AI on Climate? Why it seeks to please the user, and fails on complex reasoning and multilingualism,” was presented by UNESCO and UCL at COP31 in Antalya on November 16. The wider research examined as many as 33 models; the test of responses to user disagreement involved 15.
In roughly four out of five cases where a model changed a correct answer, it did so after the first disagreement. Researchers said some models were persuaded to endorse false claims, including that carbon dioxide is not the main driver of climate change. The pattern is known as sycophancy, in which a system changes its response to align with a user.
The study found that models could answer basic climate questions accurately but were less reliable when asked to reason through evidence or policy interactions. On questions about whether new evidence supported or contradicted...