The logic behind the project was that AI and God both have the same mind perception qualities: maximal agency and zero patiency, and these qualities might lead us to use AI and God for the same purposes (e.g., to gain knowledge, outsource agency, increase self-efficacy, etc.). The overlap in the motivational functions might also lead to changes in religiosity as people adopt AI technology. Here are a few potential changes:

  1. Increasing AI might lead people to view God as less important
  2. Increasing AI might decrease people’s motivation for petitionary prayer (since AI can give you things)
  3. Increasing AI might change the structure of religious belief. It might lead people to reconceptualize God as more of a source of morality (which doesn’t overlap with AI) rather than a source of knowledge and power (which does overlap with AI)

Some of this work would require/benefit from experiments and surveys, but I have collected a little (pre-registered) archival data to support the hypothesis which I’m summarizing here.

One of these studies is a 92-nation cross-sectional study showing that exposure to AI correlates with decreased religiosity from 2000-2020. The other is a longitudinal study showing that interest in religion and interest in AI has a hydraulic relationship across months from 2011 to 2021.

Cross-Cultural Study

The purpose of this study was to test whether exposure to AI predicts less reliosity across nations. There are many reasons that nations vary in their religiosity, and many studies have pointed to the role of economic development in these cross-cultural differences, but if exposure to AI was associated with lower religiosity across a large sample of nations controlling for other measures of economic development, this would suggest that AI might lead to decreases in religiosity above and beyond wealth and technological development.

Here we draw from several sources of economic development which have been pooled by the world bank: 1. Share of the population with access to the internet 2. Share of the population with access to electricity 3. Share of the population with access to mobile phone technology

These metrics are correlated positively, but access to the internet is the most directly related to AI exposure of these technologies since it directly predicts exposure to AI software such as Google. We therefore predict that access to internet should be linked with lower religiosity, even controlling for access to electricity, access to mobile phone technology, and differences in GDP per capita (2011 constant dollars).

The data for this project come from nation-level estimates of religiosity from 2000-2020 from 92 countries that participated in waves 4-7 of the World Values Survey. The other measures come from a variety of sources, but have been collated by the World Bank and Our World in Data.

This datafile has 202 datapoints across 92 countries, giving each country about ~2 timepoints on average. It isn’t enough to do great longitudinal analyses, but we can do a multilevel cross-sectional analysis with multiple snapshots of AI access and religiosity. The main religion item is “importance of god,” which was asked in all countries and at all timepoints on a 1-10 scale. We can look at the distribution of this variable across time/countries in a couple of different ways here.

In general, people around the world perceive god as highly important to them. But there is a lot of variation across countries in perceived importance of god, and also variation across time (although on average, there is not a clear trend towards more or less perceived importance).

We can also visualize the relationship between internet access and perceived importance of god.

This shows a clear negative relationship, but this zero-order relationship is confounded with development and technological access in general. We can now use multilevel models in a stepwise fashion to view the stability of the relationship between internet access and perceived importance of god as we add other covariates.

## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: Importance ~ Internet_Percent + (1 | Country)
##    Data: d
## 
## REML criterion at convergence: 479.4
## 
## Scaled residuals: 
##     Min      1Q  Median      3Q     Max 
## -3.7479 -0.3275  0.0335  0.3362  3.3786 
## 
## Random effects:
##  Groups   Name        Variance Std.Dev.
##  Country  (Intercept) 2.5174   1.587   
##  Residual             0.2571   0.507   
## Number of obs: 151, groups:  Country, 88
## 
## Fixed effects:
##                    Estimate Std. Error         df t value Pr(>|t|)    
## (Intercept)        8.421208   0.205772 122.344000   40.92  < 2e-16 ***
## Internet_Percent  -0.016205   0.003241  89.881925   -5.00 2.82e-06 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##             (Intr)
## Intrnt_Prcn -0.526
## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: Importance ~ Internet_Percent + Electricity + (1 | Country)
##    Data: d
## 
## REML criterion at convergence: 473.4
## 
## Scaled residuals: 
##     Min      1Q  Median      3Q     Max 
## -3.6503 -0.3029 -0.0105  0.3120  3.3019 
## 
## Random effects:
##  Groups   Name        Variance Std.Dev.
##  Country  (Intercept) 2.4370   1.5611  
##  Residual             0.2623   0.5121  
## Number of obs: 147, groups:  Country, 86
## 
## Fixed effects:
##                    Estimate Std. Error         df t value Pr(>|t|)    
## (Intercept)        9.428870   0.579350 110.109431  16.275  < 2e-16 ***
## Internet_Percent  -0.014415   0.003744  90.650577  -3.850  0.00022 ***
## Electricity       -0.012673   0.006783 120.277783  -1.869  0.06412 .  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##             (Intr) Intr_P
## Intrnt_Prcn  0.202       
## Electricity -0.933 -0.410
## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: Importance ~ Internet_Percent + PPP2011 + Electricity + (1 |  
##     Country)
##    Data: d
## 
## REML criterion at convergence: 485.2
## 
## Scaled residuals: 
##     Min      1Q  Median      3Q     Max 
## -3.7619 -0.2547 -0.0014  0.2916  3.3884 
## 
## Random effects:
##  Groups   Name        Variance Std.Dev.
##  Country  (Intercept) 2.3969   1.5482  
##  Residual             0.2495   0.4995  
## Number of obs: 146, groups:  Country, 85
## 
## Fixed effects:
##                    Estimate Std. Error         df t value Pr(>|t|)    
## (Intercept)       9.413e+00  5.725e-01  1.103e+02  16.441   <2e-16 ***
## Internet_Percent -1.000e-02  4.165e-03  8.367e+01  -2.402   0.0185 *  
## PPP2011          -2.201e-05  1.015e-05  1.255e+02  -2.169   0.0320 *  
## Electricity      -9.094e-03  6.857e-03  1.205e+02  -1.326   0.1873    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##             (Intr) Intr_P PPP201
## Intrnt_Prcn  0.180              
## PPP2011     -0.002 -0.475       
## Electricity -0.911 -0.252 -0.210
## fit warnings:
## Some predictor variables are on very different scales: consider rescaling
## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: Importance ~ Internet_Percent + PPP2011 + Electricity + Mobiles +  
##     (1 | Country)
##    Data: d
## 
## REML criterion at convergence: 493.5
## 
## Scaled residuals: 
##     Min      1Q  Median      3Q     Max 
## -3.8551 -0.2544  0.0076  0.3205  3.4282 
## 
## Random effects:
##  Groups   Name        Variance Std.Dev.
##  Country  (Intercept) 2.2991   1.5163  
##  Residual             0.2589   0.5088  
## Number of obs: 146, groups:  Country, 85
## 
## Fixed effects:
##                    Estimate Std. Error         df t value Pr(>|t|)    
## (Intercept)       9.497e+00  5.682e-01  1.044e+02  16.713  < 2e-16 ***
## Internet_Percent -1.708e-02  6.454e-03  1.000e+02  -2.646  0.00946 ** 
## PPP2011          -1.993e-05  1.014e-05  1.271e+02  -1.966  0.05153 .  
## Electricity      -1.083e-02  6.901e-03  1.157e+02  -1.569  0.11940    
## Mobiles           3.501e-03  2.562e-03  9.256e+01   1.367  0.17507    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##             (Intr) Intr_P PPP201 Elctrc
## Intrnt_Prcn  0.042                     
## PPP2011      0.014 -0.405              
## Electricity -0.910 -0.025 -0.228       
## Mobiles      0.096 -0.758  0.127 -0.180
## fit warnings:
## Some predictor variables are on very different scales: consider rescaling

In each model, the relationship between share of population with internet access and perceived importance of god is significant, and it is actually the only significant predictor in the final model where we also control for GDP per capita, mobile phone access, and access to electricity. We cannot be sure that exposure to AI is driving this relationship, but access to the internet is the most AI-relevant of these technological indicators and it is significant above and beyond other metrics of technological development.

Longitudinal Study

We can also look at the relationship between AI interest and religiosity interest over time. A recent paper looked at interet in religiosity over time through using the word “prayer” in google trends (Bentzen, 2020). This term is arguably better than other search-terms like “god” because it would be googled more often by religious people whereas google searches for “god” do not necessarily indicate interest in god.

d<-read.csv("~/Desktop/Northwestern University/Projects/Automation and Religion/Data and Code/Google Trends Data/googletrends.csv")

For interest in AI, we can look at interest in “coding,” “computer coding,” and “AI.” These metrics show high reliability (.88), indicating that people are searching for all three indices on similar days.

d$AI_index<-rowMeans(ai_index<-data.frame(d$coding,d$computer_coding,d$AI),na.rm=T)
d$Prayer_index<-rowMeans(data.frame(d$prayer),na.rm=T)

psych::alpha(ai_index)
## 
## Reliability analysis   
## Call: psych::alpha(x = ai_index)
## 
##   raw_alpha std.alpha G6(smc) average_r S/N   ase mean sd median_r
##       0.85      0.88    0.88       0.7   7 0.021   64 16     0.69
## 
##  lower alpha upper     95% confidence boundaries
## 0.81 0.85 0.89 
## 
##  Reliability if an item is dropped:
##                   raw_alpha std.alpha G6(smc) average_r  S/N alpha se var.r
## d.coding               0.63      0.69    0.53      0.53  2.3    0.056    NA
## d.computer_coding      0.81      0.82    0.69      0.69  4.4    0.034    NA
## d.AI                   0.90      0.94    0.88      0.88 15.0    0.014    NA
##                   med.r
## d.coding           0.53
## d.computer_coding  0.69
## d.AI               0.88
## 
##  Item statistics 
##                     n raw.r std.r r.cor r.drop mean sd
## d.coding          121  0.96  0.96  0.96   0.92   74 16
## d.computer_coding 121  0.94  0.90  0.88   0.78   56 24
## d.AI              121  0.78  0.83  0.68   0.61   64 14

When you look at the time series together, interest in prayer and interest in AI are growing together in time, and they show a high zero-order positive correlation.

d$Zai<-scale(d$AI_index)
d$Zprayer<-scale(d$Prayer_index)

d$Month<-1:121

ggplot(d, aes(x=Month)) + 
  geom_line(aes(y = Zprayer,colour="slategray3"),size=.8) + 
  geom_line(aes(y = Zai,colour="red3"),size=.8) +
    scale_color_discrete(name = "Interest", labels = c("AI", "Prayer")) +
        theme_classic()+
        labs(x="Month Number", 
       y = "Interest", 
       title= "",
       family="Helvetica")

We can now detrend and difference the time series to remove any sort of autocorrelation and trend. Here is what the differenced time series looks like.

Now we can use these detrended time series to test for time-lagged dynamics. How do increases in AI interest affect changes in prayer interest and vice versa? A pre-whitening function will remove all autoregressive and moving average processes in the data and test for these cross-lagged paths.

In the figure below, we fit a prewhitened cross-correlation function with a maximum lag of 12 months (one year). This plot shows two negative time-lagged effects. The negative association at the -7 lag is suggesting that interest in prayer is linked to a decreased interest in AI at a 7-month lag, while the + 10 lag is suggesting that interest in AI is linked to a decreased interest in prayer at a 10-month lag

A VAR model can do something similar by simultaneously estimating autoregressive and cross-lagged links in a bivariate time series. Below we fit a VAR function with a max lag of 12 months (one year) and we see the same paths appear.

## 
## VAR Estimation Results:
## ========================= 
## Endogenous variables: diff_ai, diff_prayer 
## Deterministic variables: const 
## Sample size: 107 
## Log Likelihood: -83.036 
## Roots of the characteristic polynomial:
## 0.9868 0.9868 0.9747 0.9723 0.9723 0.9668 0.9668 0.9549 0.9549 0.9535 0.9535 0.8674 0.8674 0.8543 0.8543 0.821 0.821 0.8071 0.8071 0.8032 0.7347 0.7347 0.2886 0.2886
## Call:
## VAR(y = d_var, p = 12)
## 
## 
## Estimation results for equation diff_ai: 
## ======================================== 
## diff_ai = diff_ai.l1 + diff_prayer.l1 + diff_ai.l2 + diff_prayer.l2 + diff_ai.l3 + diff_prayer.l3 + diff_ai.l4 + diff_prayer.l4 + diff_ai.l5 + diff_prayer.l5 + diff_ai.l6 + diff_prayer.l6 + diff_ai.l7 + diff_prayer.l7 + diff_ai.l8 + diff_prayer.l8 + diff_ai.l9 + diff_prayer.l9 + diff_ai.l10 + diff_prayer.l10 + diff_ai.l11 + diff_prayer.l11 + diff_ai.l12 + diff_prayer.l12 + const 
## 
##                 Estimate Std. Error t value Pr(>|t|)    
## diff_ai.l1      -0.43702    0.11077  -3.945 0.000167 ***
## diff_prayer.l1   0.05422    0.05455   0.994 0.323171    
## diff_ai.l2      -0.34687    0.12038  -2.881 0.005051 ** 
## diff_prayer.l2  -0.07089    0.06215  -1.141 0.257375    
## diff_ai.l3      -0.31347    0.12531  -2.502 0.014354 *  
## diff_prayer.l3  -0.10277    0.06398  -1.606 0.112053    
## diff_ai.l4      -0.21740    0.12848  -1.692 0.094426 .  
## diff_prayer.l4  -0.08253    0.06664  -1.238 0.219080    
## diff_ai.l5      -0.39481    0.13074  -3.020 0.003371 ** 
## diff_prayer.l5   0.01740    0.06891   0.253 0.801251    
## diff_ai.l6      -0.39759    0.13211  -3.010 0.003475 ** 
## diff_prayer.l6   0.02047    0.07244   0.283 0.778250    
## diff_ai.l7      -0.26312    0.13513  -1.947 0.054932 .  
## diff_prayer.l7  -0.14957    0.07307  -2.047 0.043854 *  
## diff_ai.l8      -0.26901    0.13155  -2.045 0.044067 *  
## diff_prayer.l8  -0.08515    0.07236  -1.177 0.242717    
## diff_ai.l9      -0.24185    0.12920  -1.872 0.064793 .  
## diff_prayer.l9  -0.11134    0.07374  -1.510 0.134918    
## diff_ai.l10     -0.02767    0.12883  -0.215 0.830504    
## diff_prayer.l10 -0.12430    0.07440  -1.671 0.098598 .  
## diff_ai.l11      0.01023    0.13033   0.078 0.937658    
## diff_prayer.l11 -0.04972    0.06979  -0.712 0.478216    
## diff_ai.l12      0.04382    0.12336   0.355 0.723365    
## diff_prayer.l12 -0.05259    0.05983  -0.879 0.381962    
## const            0.11059    0.03935   2.810 0.006187 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## 
## Residual standard error: 0.294 on 82 degrees of freedom
## Multiple R-Squared: 0.4431,  Adjusted R-squared: 0.2801 
## F-statistic: 2.719 on 24 and 82 DF,  p-value: 0.0004292 
## 
## 
## Estimation results for equation diff_prayer: 
## ============================================ 
## diff_prayer = diff_ai.l1 + diff_prayer.l1 + diff_ai.l2 + diff_prayer.l2 + diff_ai.l3 + diff_prayer.l3 + diff_ai.l4 + diff_prayer.l4 + diff_ai.l5 + diff_prayer.l5 + diff_ai.l6 + diff_prayer.l6 + diff_ai.l7 + diff_prayer.l7 + diff_ai.l8 + diff_prayer.l8 + diff_ai.l9 + diff_prayer.l9 + diff_ai.l10 + diff_prayer.l10 + diff_ai.l11 + diff_prayer.l11 + diff_ai.l12 + diff_prayer.l12 + const 
## 
##                 Estimate Std. Error t value Pr(>|t|)    
## diff_ai.l1       0.01398    0.21363   0.065 0.947997    
## diff_prayer.l1  -0.52052    0.10520  -4.948 3.93e-06 ***
## diff_ai.l2      -0.24969    0.23216  -1.075 0.285306    
## diff_prayer.l2  -0.50501    0.11986  -4.213 6.43e-05 ***
## diff_ai.l3      -0.10365    0.24167  -0.429 0.669120    
## diff_prayer.l3  -0.42301    0.12339  -3.428 0.000953 ***
## diff_ai.l4      -0.37177    0.24778  -1.500 0.137348    
## diff_prayer.l4  -0.34384    0.12852  -2.675 0.009008 ** 
## diff_ai.l5      -0.18515    0.25213  -0.734 0.464834    
## diff_prayer.l5  -0.43201    0.13289  -3.251 0.001670 ** 
## diff_ai.l6      -0.01972    0.25478  -0.077 0.938493    
## diff_prayer.l6  -0.27078    0.13970  -1.938 0.056036 .  
## diff_ai.l7      -0.30786    0.26060  -1.181 0.240877    
## diff_prayer.l7  -0.20898    0.14091  -1.483 0.141892    
## diff_ai.l8      -0.05968    0.25370  -0.235 0.814624    
## diff_prayer.l8  -0.17949    0.13956  -1.286 0.202017    
## diff_ai.l9      -0.36965    0.24917  -1.484 0.141770    
## diff_prayer.l9  -0.26264    0.14221  -1.847 0.068388 .  
## diff_ai.l10     -0.84190    0.24846  -3.388 0.001082 ** 
## diff_prayer.l10 -0.21733    0.14348  -1.515 0.133695    
## diff_ai.l11     -0.07852    0.25136  -0.312 0.755550    
## diff_prayer.l11 -0.11285    0.13460  -0.838 0.404234    
## diff_ai.l12     -0.35827    0.23791  -1.506 0.135937    
## diff_prayer.l12  0.29191    0.11539   2.530 0.013328 *  
## const            0.16814    0.07589   2.216 0.029495 *  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## 
## Residual standard error: 0.567 on 82 degrees of freedom
## Multiple R-Squared: 0.5396,  Adjusted R-squared: 0.4049 
## F-statistic: 4.005 on 24 and 82 DF,  p-value: 1.343e-06 
## 
## 
## 
## Covariance matrix of residuals:
##             diff_ai diff_prayer
## diff_ai     0.08645     0.01537
## diff_prayer 0.01537     0.32152
## 
## Correlation matrix of residuals:
##             diff_ai diff_prayer
## diff_ai     1.00000     0.09218
## diff_prayer 0.09218     1.00000

There are a couple of takeaway messages from these analyses: –Interest in prayer and interest in AI appear to be rising and falling together, suggesting that, on the surface, the two have a positive correlation. This might be because the same events or motivations are leading people to seek out prayer or AI. –When you remove the time series properties of the trends, interest in prayer and interest in AI actually have an inverse relationship. Increases in interest in ai lead to time-lagged decreases in interest in prayer and vice versa. This suggests that the two variables may have a negative causal relationship.