Exposure to automation explains religious declines

Our first study tracked religious declines across world nations. We operationalized automation through each nation’s yearly operational stock of industrial robots, defined as an “automatically controlled, reprogrammable multipurpose manipulator programable in three or more axes” by the International Federation of Robots (IFR). We operationalized religiosity through yearly survey data on the proportion of people across nations ( $\sum n$ = 2,014,633) who answered “yes” to the question “Is religion an important part of your daily life?” Variations of this religious importance item are frequently used to measure religiosity across cultures because it can gauge religiosity across a variety of religious traditions and does not make assumptions about religious content (e.g., monotheism). Combining these datasets gave us data on 68 countries from 2006 to 2019, which we used to test whether robotics exposure could explain 21st-century religious decline.

We controlled for several other variables. Our primary control variables involved other forms of technological change: mobile phone subscriptions per capita and the share of population with access to electricity. Measuring telecommunication and energy development allowed us to test the role of robotics above and beyond general technological infrastructure. We also controlled for population size, since operational stock of robots could be larger in more populous countries. We also controlled for GDP per capita and individual choice norms around fertility. We controlled for these latter two variables because the existential theory of secularization suggests that wealth reduces religiosity because it increases certainty and stability (25), whereas other theories focus on value change (32); one view suggests that religious decline occurs when cultures emphasize individual choice norms (e.g., contraception) over profertility norms (27). Our Methods section contains more information about each of these covariates.

Cross-sectional models with intercepts varying randomly across nations found that robotics exposure was robustly and negatively associated with religiosity across the globe (Table 1, model 1). This negative association replicated controlling for GDP per capita and population size (Table 1, model 2) and continued to reach significance controlling for telecommunication and energy development (Table 1, model 3). A longitudinal model with intercepts and slopes randomly varying across nations next estimated whether robotics was linked to declines in religiosity. In this model (Table 1, model 4), the interaction between robotics and year was significant. This effect replicated controlling for the interaction of telecommunication development with year (Table 1, model 5) and energy development with year (Table 1, model 6). Neither energy nor telecommunication development significantly explained declines in religiosity. In the model where we held these covariates constant, nations with a high operational stock (+1 SD) of robots experienced an approximately 3% decline in religiosity per decade (P = 0.01), whereas nations with low operational stock (−1 SD) showed approximately a 0.1% increase per decade (P = 0.95). This may seem like a small effect, but it was substantially larger than any other geopolitical variable, and these small effects can fuel large divergences in religiosity over time. Fig. 1 illustrates these dynamics.

These cross-cultural analyses show that robotics exposure, measured here through the density of industrial robots, can explain variation in religiosity around the world and variation in global religious decline from 2006 to 2019. Robotics exposure was associated with religious decline above and beyond other forms of technological development, such as telecommunications development and energy development. Each of these results held controlling for GDP per capita and population size.

In our supplemental analyses, we show that key findings replicate when controlling for spatial autocorrelation and when removing majority-Muslim countries, which had low rates of robot workers and also low rates of religious decline. Analyses also replicate when we interact all variables with year instead of just technological innovations and when we control for an alternative measure of individualism which is more outdated but includes more nations. One interesting finding that emerges in this supplemental analysis is that individual choice norms around fertility, which have a strong negative cross-sectional relationship with religiosity (Table 1), do not significantly predict 21st-century religious decline. We discuss this curious result more in our general discussion.

Study 2 next tested for the relationship between robotics exposure and religious decline within a single nation—the United States of America (USA). Testing the relationship within the USA offered a more conservative test of our hypothesis, since regions within the USA are more religiously homogenous than world nations and have more similar levels of technological development. We therefore could test whether robotics exposure explains religious decline controlling for other variables, such as income, employment rate, and residential mobility, which do vary considerably across regions of the USA and have been linked to automation (33, 34).

We again measured religiosity through the self-reported importance of religion—measured from 2008 to 2016 across metropolitan areas—and measured robotics exposure through the operational stock of industrial robots. Specifically, we used estimates released from the Brookings Institute of each metropolitan areas’ percent growth in industrial robot operational stock (henceforth called robotics growth) between 2010 and 2015. We combined these metrics into a dataset which also included estimates of median income, unemployment, and residential mobility (reverse coded as number of nonmovers) across metropolitan areas. This dataset contained estimates from 110 metropolitan areas and over 1 million individuals (Materials and Methods) within the USA, with time-varying religion data from 2008 to 2016 and time-invariant data on robotics growth, income, unemployment, and residential mobility. See Materials and Methods for more information about all variables. SI Appendix, Fig. S2 is a map of the metropolitan areas in our analysis organized by religious decline and robotics growth.

Cross-sectional models with intercepts varying randomly across metropolitan areas and states found that metropolitan areas with high rates of robotics growth showed no significant differences in religiosity compared to metropolitan areas with low rates of robotics growth (Table 2, model 1), and this association remained null after controlling for unemployment, median income, population size, and number of nonmovers (Table 2, model 2). This cross-sectional association between robotics growth and religiosity is less meaningful across metropolitan areas than across world nations because the distribution of robot workers in the USA has strong regional constraints based on the availability of warehouse space, unionization density, and proximity to travel networks (35). Our critical test was therefore whether robotics growth would explain religious declines. In this test of change over time—a longitudinal model with random intercepts and slopes—we found that robotics growth interacted with time and significantly explained religious declines (Table 2, model 3). Metropolitan areas with higher levels of robotics growth (+1 SD) experienced an approximately 3% yearly decline in religion each decade (P = 0.006)—mirroring the effect size we observed in study 1—while metropolitan areas with lower levels of robotics growth (−1 SD) experienced approximately a 0.5% yearly rise in religion (P = 0.67). This effect replicated controlling for the interaction of income and year (Table 2, model 4), population size and year, and number of nonmovers and year (Table 2, model 5). No other factor explained religious decline in these models.

These analyses show that robotics exposure can explain religious declines within a nation as well as across nations. Robotics growth was associated with religious declines across USA metropolitan areas controlling for changes in income, residential mobility, and unemployment. In our SI Appendix, we present additional analyses where the timespan of religious change is restricted to the same time window in which we have available data on robotics growth (2010 to 2015).

In sum, studies 1 to 2 showed that robotics exposure explained religious decline across and within nations. However, research on the ecological fallacy (36) and Simpson’s paradox (37) shows that group-level associations sometimes do not replicate, and can even reverse, at the individual level. Studies 1 to 2 also focused on exposure to automation through robotics, but this exposure can also happen through occupational work with AI algorithms. Studies 3 to 4 addressed both of these limitations to show that occupational exposure to AI was associated with declines in religiosity across individuals.

Study 3 was a preregistered analysis of occupational AI exposure and religiosity within an 11-wave longitudinal study, which was conducted between 2009 and 2020 in a community sample. Participants in this multiwave study answered several questions about their personal characteristics and their social attitudes. Nine waves of the study included binary items asking people if they believed in God and if they identified as religious or nonreligious (key religion items were omitted from waves 1 and 6). Participants also free-reported their occupation (if any) in all waves, which was manually coded by research assistants into one of 1,036 categories (e.g., “social security assessor” and “debt collector”). In total, 69,021 individuals participated in at least one wave of the study and 46,680 individuals reported their religion in multiple waves of the study (see Materials and Methods and SI Appendix for more information about sample and recruiting). Our critical hypothesis was that occupational AI exposure would be associated with lower religiosity across individuals and religious decline within individuals.

We measured occupational AI exposure by incorporating occupation-level metadata from O*Net, a large occupational database which has classified occupations based on the importance of different occupational qualities (see Materials and Methods for more information). We used data on the importance of programming as a proxy for AI exposure. This proxy appeared face valid since many of the jobs with high importance of programming also involve AI exposure (e.g., “Software Engineer” and “Web Developer”). We also controlled for importance of biology, chemistry, mathematics, and medicine/dentistry to ensure that generalized scientific exposure or scientific education did not confound AI exposure. We also controlled for SES, age, gender, and political conservatism (38). Our main analyses focused on God belief because it allowed us to test whether AI exposure was associated with religious beliefs rather than just self-reported importance of religiosity (which we had found in studies 1 to 2). We found similar results using the religious identification variable, and we summarize those results in SI Appendix. In general, God belief was stable across the study—God belief at wave 1 correlated at 0.75 with God belief at wave 11—but a notable proportion of people (17.39%) changed their belief at least once across waves.

Baseline analyses of occupational AI exposure and God belief showed that a random slopes and intercepts model outperformed a random intercepts model, ${\chi }^2$ = 34.42, P < 0.001. In this random slopes and intercepts model, AI exposure and religiosity were negatively and significantly associated, b = −0.59, SE = 0.06, OR = 0.55, t = −9.33, P < 0.001, 95% CIs [−0.72, −0.47]. Since occupational AI exposure was standardized through z-scoring, the odds ratio suggested that people with jobs that were one SD higher than the mean on occupational exposure to AI were 45% less likely to believe in God compared to people in occupations that had a mean level of exposure to AI. Subsequent models found that this association remained statistically significant controlling for SES, age, gender, and political conservatism (Table 3, model 1), and after controlling for generalized scientific exposure (Table 3, model 2). In subsequent models, we added lagged terms representing occupational exposure to AI at previous timepoints in the survey. In this model, the second-order lag was significantly negatively associated with God belief above and beyond the contemporaneous effect (Table 3, model 3; see SI Appendix for models including higher-order lags). Occupational exposure to AI may increase someone’s likelihood of losing their belief in God, even if they subsequently move into an occupation that no longer involves AI exposure. Fig. 2 visualizes these effects.

We also fit models which used a within-person centering procedure to separately estimate the within-person and between-person associations between occupational AI exposure and God belief (39). In order for these models to converge, we needed to restrict our sample to individuals who participated in at least six waves of the survey (n = 5,542). Results showed that occupational AI exposure was negatively associated with God belief at both the between-person level, b = −0.08, SE = 0.01, OR = 0.93, t = −6.74, P < 0.001, 95% CIs [−0.10, −0.05], and the within-person level, b = −0.02, SE = 0.008, OR = 0.98, t = −2.43, P = 0.02, 95% CIs [−0.03, −0.004]. After controlling for the generalized scientific exposure proxies, these between-person, b = −0.07, SE = 0.01, OR = 0.93, t = −6.59, P < 0.001, 95% CIs [−0.10, −0.05], and within-person, b = −0.02, SE = 0.008, OR = 0.98, t = −2.18, P = 0.03, 95% CIs [−0.03, −0.002], relationships remained negative and statistically significant. In other words, occupational AI exposure explained variation in religious belief across individuals, but also religious decline in the same individual over time.

Our SI Appendix contain several supporting analyses. These include a replication of our findings using the religious identification as an outcome instead of God belief, robustness checks which replicate our results using different subgroups of participants to rule out the possibility that our findings were confounded with selective attrition, additional descriptive statistics concerning our key variables, additional models which control for general education rather than specific scientific knowledge, and a discussion of alternative modeling procedures. All analyses support our general conclusion that occupational exposure to AI is associated with lower religiosity between individuals and religious declines within individuals.

Study 4 was a preregistered three-wave (time-lagged) study in which we measured occupational AI exposure in a manufacturing plant as it integrated AI technology. We followed 238 employees within the organization over time, directly measuring their exposure to AI and their religious profiles (intrinsic religiosity and fundamentalism) at time 1 (T1), their perceived religious importance over the last week and their frequency of religious behaviors over the last week at time 2 (T2), and supervisor-reports of employee workplace behaviors at time 3 (T3). Study 4 was the most conservative test of our hypothesis because it featured the smallest and most homogenous sample (participants were primarily Muslim; see Materials and Methods and SI Appendix) and took place over the narrowest time frame (3 wk). This study also allowed us to explore the consequences of AI-linked religious decline for workplace behavior.

Because of space limitations, we present the T1 and T2 findings in this paper, which replicate our findings in study 3. Initial analyses found that T1 AI exposure was negatively associated with T2 religiosity, r(236) = −0.24, P < 0.001. This association was stronger for the items measuring participants’ subjective religious importance over the last week, r(236) = −0.27, P < 0.001, than for the items measuring frequency of religious behaviors over the last week, r(236) = −0.17, P < 0.001. This association alone is limited because less religious people may have been faster to adopt AI technology. However, we found that the negative association replicated when we controlled for T1 religious fundamentalism and intrinsic religiosity, b = −0.18, SE = 0.08, $\beta$ = −0.19, t(229) = −2.17, P = 0.03, 95% CIs [−0.35, −0.02], suggesting that AI predicts declines in religiosity as well as cross-sectional variation in religiosity.

We provide more analyses and statistics in our SI Appendix. These analyses also consider downstream associations with participants’ T3 workplace behaviors that have been linked to religiosity (40–42) (e.g., unethical behavior, trust, incivility, and organizational citizenship behavior, which are face-valid indicators of prosociality). We find that declines in religiosity predict changes in many of these variables.

In sum, studies 3 to 4 showed that exposure to AI is linked to declines in religiosity at the individual level. Occupational exposure to AI correlated with lower levels of religiosity across and within individuals, even controlling for covariates. Moreover, AI exposure predicted future declines in religion.

Our final study was a preregistered experiment testing whether learning about advances in AI would temporarily decrease religious conviction to a greater extent than learning about other scientific advances. This experiment also explored the properties of AI that might explain why learning about AI reduces religious conviction. We were particularly interested in whether people perceive AI, like God, to operate outside the laws of nature compared to traditional sciences such as chemistry, biology, and medicine. To do this, we asked people whether they associated AI and other scientific advances with discovering and applying “laws of nature,” with the implicit view that domains associated with laws of nature like gravity, matter, and motion would be more constrained by these laws. This approach has limitations (e.g., being associated vs. constrained by laws of nature are not interchangeable), but we felt that this approach was less demand-laden than nudging participants to think of automation vs. science as God-like, which could interfere with our central dependent variable, religious conviction.

We used a between-subjects design to test our hypotheses. A general sample of 1,400 participants with a range of religious beliefs (Materials and Methods)—were randomly assigned to read about three advances in science or AI, which were matched on domain (e.g., language, medicine, or agriculture) and length (one paragraph). For example, participants in the AI condition read about ChatGPT, whereas participants in the science condition read about a recently published study showing that Broca’s area is involved in the production of sign language. Participants rated each automation/scientific advance on its impressiveness, technological sophistication, and the extent that it was associated with laws of nature. They also rated the extent that each advance increased vs. decreased their religious conviction (see Materials and Methods for more details).

We fit general linear models with Gaussian estimation testing whether AI advances decreased religious conviction relative to science advances, controlling for their impressiveness and technological sophistication. AI advances were rated as similarly impressive to science advances, b = 0.04, SE = 0.06, t = 0.66, P = 0.51, 95% CIs [−0.08, 0.17], and more technologically sophisticated than scientific advances, b = 1.26, SE = 0.07, t = 18.27, P < 0.001, 95% CIs [1.13, 1.40]. Our results are virtually identical regardless of these controls (SI Appendix).

As we predicted, participants viewed AI advances as less associated from laws of nature than scientific advances, b = −1.56, SE = 0.08, t = −20.34, P < 0.001, 95% CIs [−1.71, −1.41]. We also found that participants reported less religious conviction in the AI condition vs. the science condition, b = −0.71, SE = 0.10, t = −6.95, P < 0.001, 95% CIs [−0.91, −0.51]. The effect on religious conviction was larger among participants who identified as religious, b = −1.05, SE = 0.14, t = −7.36, P < 0.001, 95% CIs [−1.34, −0.77], than nonreligious, b = −0.19, SE = 0.10, t = −1.85, P = 0.06, 95% CIs [−0.39, 0.01], presumably because religious conviction had a greater range of variance for religious participants. Effects broken down by each domain of innovation are summarized in Fig. 3. Fig. 3 also reports a 5,000-sample bootstrapped mediational model in which find that the association with laws of nature fully mediated the effect of automation on religious conviction.

Studies 1 to 2 did not involve original human subject data collection. Our original human subject studies were approved by Institutional Review Boards. Study 3 was approved by the University of Auckland’s Human Ethics Committee, with reviews and renewals every 3 y (the most recent reference number is #014889). Study 4 was approved by National Sun Yat-Sen University’s IRB (IRB# 200604). Study 5 was approved by Northwestern University’s IRB (STU# STU00206763).

J.C.J. and A.W. designed research; J.C.J., K.C.Y., P.M.T., and C.G.S. performed research; C.G.S. contributed new reagents/analytic tools; J.C.J. and A.W. analyzed data; and J.C.J. and A.W. wrote the paper.

The authors declare no competing interest.