arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

ChatGPT‑like Models Can Guide Voters, But Carry Noticeable Biases | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
cs.CYEnglishPublished

ChatGPT‑like Models Can Guide Voters, But Carry Noticeable Biases

November 20, 2026arXiv: 2611.24684v1

A recent study tested a commercial large language model (LLM) on the task of voting advice. Researchers uploaded policy preferences from 6,022 German voters to Gemini 3 Flash and asked the LLM to recommend a party for the 2025 federal election. The model output a party name for each voter.

Comparing the LLM’s advice to the voter’s actual choice and to a traditional salience‑weighted proximity model (the “Expert model”), the study found that in about 70 % of cases the LLM suggested the same party as the Expert model. The model also tipped the balance toward the far‑right Alternative for Germany (AfD) at 29 %, roughly 1.5 times the 19 % that the Expert model would recommend, and it rarely advised voting for Die Linke (2 % vs 5 %) or the BSW.

The biggest gain came from voters with low political interest or education. For these respondents, the LLM’s advice improved alignment with their policy preferences more than the Expert model would have. Still, the study warns that the LLM’s voting suggestions display systematic biases that are hard to explain.

Overall, off‑the‑shelf LLMs can emulate a basic voting‑advice application, yet users should be cautious about the LLM’s tendency to over‑recommend the AfD and under‑recommend smaller parties.

Source paper

Hi Gemini, Who Should I Vote For? The Benefits (and Problems) of Using LLMs As Voting Advice Applications

cs.CY
View on arXiv