Reply to Moore et al.: Manipulation adherence and baseline AI attitudes as moderators of the effect of God salience on algorithm aversion

We found that God salience leads to lower algorithm aversion in several preregistered experiments, supplemented by secondary data (1). Moore et al. (2), however, found weaker results when replicating our online studies. While the results were significant in one study, they were merely directional in others and opposite of predictions in one. We compared our findings to those of Moore et al. and identified two areas for discussion and future research.

First, as noted by Moore et al. (3), the effect of the manipulation on the God salience manipulation check was smaller in their direct replication of study 1, sparking questions regarding participants’ adherence to the manipulations. In a reanalysis of Moore et al.’s study 1, we found a marginally significant interaction of God salience by God salience manipulation check responses (β = −1.65, P < 0.06); when participants in the God salient condition actually reported lower levels of thinking about God, they were significantly less positive (M = 52.12) toward AI-based recommendations than those who reported higher levels of thinking about God (M = 55.98; β = −1.52, P = 0.006). This dampened the means in the God condition but is broadly consistent with our hypothesis regarding God salience. The same pattern holds in our study 1, wherein the effect within God condition is weaker among those who reported lower (M = 51.54), as opposed to higher (M = 48.10), levels of God salience (β = −1.67, P = 0.03). Even in Moore et al.’s conceptual replication of study 1 (where the overall effect was in the opposite direction of predictions), the pattern is in the predicted direction among participants who reported thinking about God to a larger extent (M_control = 55.22, M_God = 51.42; β = −3.80, P = 0.08). Unfortunately, we do not have similar manipulation checks in the remaining studies to further explore this pattern. Furthermore, this does not explain why adherence to the manipulation differs. Greater statistical power and a more nuanced perspective on the drivers of different levels of God salience adherence will be helpful moving forward.

Second, we show that God salience lessens reliance on humans because people become indifferent between human and AI-based recommendations. However, if people are becoming indifferent at baseline given increased experience with algorithms (4–6), the effect of God salience should become weaker. Greater indifference may be particularly likely as people are now experiencing generative AI and its impressive ability to imitate humans (7, 8). We calculated the percentage of people choosing human (versus AI recommendations) across the binary choice studies that Moore et al. attempted to replicate (i.e., original studies 2a, 2b, and 3) and found that 60% chose humans. In our original studies, conducted beginning in 2021, people preferred human recommendations more often (68% in the subset replicated by Moore et al., 71% in total). This suggests that a shift in AI attitudes is worth exploring.

We appreciate the opportunity for this dialogue and look forward to better understanding drivers of God salience and evolving attitudes toward AI.