In recent years, large language models have influenced several activities that users do on the Internet and have enabled the automation of many customers support services. As expected, large language models gain their knowledge from the content used to train them. Given that large language models trained in different countries and with content in different languages are available, some authors have observed that large language models copy the values and prejudices encountered in those countries. In this paper, our goal is to compare attitudes about corruption obtained by surveying citizens in 59 countries worldwide with those obtained by four large language models. The large language models used were created in the United States, China, Brazil and Israel, and the assumption is that they reflect the attitudes of the nation that created the content on which individual models were trained. The Manhattan distance measurement method in 9-dimensional space was used to compare the similarities, given that the survey covering corruption consists of 9 Likert scale questions from the 2022 World Wave Survey. The results indicate that large language models’ inclination towards corruption is partially culturally biased.