Poster #81 - Themes of Deconversion Using GPT-3

• In Event: Poster Session 01

Abstract

Over the last two decades the percentage of people who are religiously unaffiliated has more than doubled, and often the exit happens during adolescence and young adulthood (Hardy & Longo, 2019). This is concerning, as religiousness bolsters positive outcomes (e.g., health and morality) and safeguards against negative outcomes (e.g., substance use and depression; Hardy et al., 2019). However, there is limited research on religious deconversion and most of what is known is based on informal research regarding self-reported reasons for leaving religion (Hardy & Longo, 2019). The purpose of the present study was to harness the power of the internet to more systematically examine reasons youth leave religion.

Data for the present study were generated by the GPT-3 (Generative Pre-trained Transformer 3) language model, which uses artificial intelligence and billions of parameters to generate text based on information available online. To compare the data and results to that from human subjects research, we used findings from the National Study of Youth and Religion (NSYR). Their sample included 250 youth who grew up religious but were no longer religiously affiliated. They were asked for the reasons why they had left religion, and those reasons were coded and ranked. We pulled demographic characteristics for those 250 youth. That set of demographic characteristics was used to prompt the GPT-3 model. Essentially, we simulated conversations between youth with those demographics and an interviewer who asked them their reasons for leaving religion. Then, GPT-3 generated text answering those questions using predictive probabilities. In essence, this procedure generated 250 vignettes about reasons for deconversion based on reasons people have posted online. A group of four human coders then coded these vignettes using open coding. Each coder grouped open codes into main themes; main themes were then compared among coders and grouped into core themes. Lastly, we compared our core themes to themes identified in the NSYR dataset.

The following are the top deconversion reasons identified: intellectual skepticism, expanding perspective/education, religion is irrational/harms, religious change outcomes (i.e., identity shifts), gradual deconversion process, and stopped attending (see Table 1). Three of our core themes matched those from the NSYR. These themes were stopped attending (NSYR called it “just stopped attending”), religion is irrational/harms (NSYR called it “disliked religion”), and intellectual skepticism (NSYR called it “intellectual skepticism and disbelief”). Moreover, the frequency of occurrence for some of our replicated core themes is comparable to the frequency of the NSYR themes.

This partial replication supports the validity and utility of GPT-3 generated data for social sciences research. In this case, it provided a systematic approach to understanding contemporary reasons for religious deconversion among youth. Future work will attempt to run experiments using GPT-3 to manipulate various individual and contextual factors to examine their roles in the deconversion process.

Authors

• Elizabeth A Nielsen, Brigham Young University
  Presenting Author

• Emelia Call-Tadje, Brigham Young University

• Alejandro Rodriguez, Brigham Young University

• Emily Marie Taylor, Brigham Young University

• Jenae M Nelson, Baylor University

• David Wingate, Brigham Young University

• Sam A Hardy, Brigham Young University