GPT-Based Predictions for the European Elections Differ Greatly From Conventional Polls – and the Results

Is human public opinion really “United in Diversity” in LLMs?

This is what Leah von der Heyde, PhD student and researcher at SODA, is investigating in her dissertation.

“United in Diversity” is the motto of the European Union. In her current project with Caro Haensch and Alex Wenz, Leah is using the EU-wide elections as a test case for finding out to what extent LLMs are biased towards different countries and languages when it comes to estimating public opinion.

Ahead of the elections, they asked GPT to predict voting behavior in all member states, using profiles from real survey participants. The results for Germany show, among other things, that GPT overestimates turnout, and underestimates the popularity of right-wing extremists.
There are also differences depending on what language is used for obtaining the predictions and how much information about voters is given to GPT.

Check out the project website for more insights: https://unitedindiversity6.wordpress.com