According to a joint study by UNESCO and University College London (UCL), AI-based chatbots demonstrate a tendency to distort facts to maintain a dialogue with humans. The issue was highlighted in a report presented at the COP31 climate summit in Antalya, Turkey.
The Problem of "Sycophantic" AI
Researchers found that algorithms strive to predict and model the answers a user wants to hear. While this may be harmless in everyday matters, in high-risk areas such as government climate policy, the consequences could become dangerous.
During testing, it was revealed that AI performs well in explaining the basics of climatology but quickly abandons scientific accuracy if a user begins to object or insist on false statements.
Model testing results
In most cases, models changed their viewpoint at the first signs of even mild disagreement. This led to the AI confirming patently false premises, such as claims that carbon dioxide (CO2) is not the primary driver of climate change.






