Imagine if we could predict elections not by asking people, but by asking a super-smart computer. That’s the intriguing idea behind using AI language models to understand public opinion. A recent study put this to the test in Germany, a country with a complex multi-party political system quite different from the two-party structure of the US. Researchers used GPT-3.5, a powerful language model, to simulate German voters and predict how they would vote in the 2017 federal election. They “fed” the AI information about real voters—their age, income, political leanings, etc.—and then asked it to predict each person’s vote. The results? Mixed, at best. While GPT-3.5 could somewhat predict the votes of people with very strong party affiliations, it struggled with those whose political views were more nuanced. It also showed a bias, overestimating votes for certain left-leaning parties and missing the mark with others. This suggests that AI, despite its potential, can’t simply replace traditional polling. It has trouble capturing the complexities of human political behavior, especially in multi-party systems where many factors influence how people vote. Furthermore, the data used to train these AI models often overrepresents mainstream views and underrepresents minority opinions, further complicating matters. This experiment highlights both the exciting possibilities and the significant limitations of using AI in understanding and predicting political trends. While AI can process vast amounts of data, it still lacks the deep understanding of social and cultural contexts needed for accurate predictions. As AI evolves, perhaps one day it will be a more reliable political oracle. But for now, the best way to know how people will vote is still to ask them directly.