Research on human–AI decision-making reports opposite baseline findings: people sometimes prefer algorithmic to human judgment (algorithm appreciation) and sometimes the reverse (algorithm aversion). A 2025 meta-analysis of 442 effect sizes from 163 studies (N = 82,078) proposed the capability–personalization framework — AI appreciation (d=0.27d = 0.27) when AI is perceived as more capable than humans and personalization is perceived as unnecessary in the decision context; AI aversion (d=−0.50d = -0.50) otherwise — but its support is observational, and its moderation analyses reveal substantial unexplained heterogeneity (tangible robots vs. intangible algorithms, attitudinal vs. behavioral outcomes, between- vs. within-subjects designs, country-level education, internet penetration, and unemployment). A 2024 MIS Quarterly research note additionally shows that the aversion and appreciation constructs lack shared definitions and measures, so the observed direction of preference may depend on which comparison a study measures. The open problem is to determine and causally validate the specific conditions under which the direction of preference between AI/algorithmic and human judgment reverses: experimentally testing the causal role of perceived relative capability and perceived personalization necessity, adjudicating whether the reported moderators are genuine causal boundary conditions or methodological artifacts, and establishing measurement configurations under which preference direction is comparable across studies.

Background

People increasingly receive advice, recommendations, and decisions from artificial intelligence (AI) and algorithms — formal computational procedures, software systems, or robots — in domains from entertainment recommendations to medical triage. Yet studies of whether people prefer algorithmic to human judgment point in opposite directions. Dietvorst, Simmons, and Massey (2015) documented "algorithm aversion": after seeing an algorithm err, people abandon it even when it predicts better than a human judge. Logg, Minson, and Moore (2019) reported the opposite, "algorithm appreciation": laypeople weight advice from an algorithm more heavily than identical advice from another person. Longoni, Bonezzi, and Morewedge (2019) showed domain-contingent resistance in medicine: consumers reject medical AI unless it substantially outperforms human physicians. Subsequent primary studies accumulated conflicting results about which direction obtains when.

Qin, Zhou, Chen, Wu, Zhou, Dong, Cao, and Lu (2025) reconciled this literature through a meta-analysis — a quantitative synthesis of results across prior studies — of 442 effect sizes from 163 studies (N = 82,078), and proposed the capability–personalization framework: when choosing between an AI and a human for a decision, people weigh two dimensions, (a) AI's perceived capability relative to humans and (b) the perceived necessity of personalization, i.e., whether the decision is seen as requiring uniquely human understanding, empathy, or tailoring to the individual. Meta-analytically, when AI is perceived as more capable than humans and personalization is perceived as unnecessary, AI appreciation emerges (d=0.27d = 0.27, 95% CI [0.17, 0.37], where dd is the standardized mean difference in preference for AI over humans); otherwise AI aversion emerges (d=−0.50d = -0.50, 95% CI [-0.63, -0.37]). Moderation analyses — tests of whether the average effect differs across study characteristics — nevertheless revealed substantial heterogeneity: appreciation is more pronounced for tangible robots than intangible algorithms, for attitudinal than behavioral outcome measures, in between-subjects designs (each participant evaluates only one advice source) than within-subjects designs (each participant compares sources), and in low-unemployment countries, while aversion is more pronounced in countries with high education and internet penetration. Because this evidence pools existing studies rather than manipulating the two framework dimensions, it cannot establish that perceived capability and personalization necessity causally control the sign of the preference, nor whether the moderating factors are causal boundary conditions or artifacts of design and measurement.

Independently, Jussupow, Benbasat, and Heinzl (2024) argued that the aversion/appreciation literature lacks a shared understanding of the two constructs' meanings and measurements: studies operationalize "preference" as different comparisons — of algorithms against human advisors, against one's own judgment, or against combinations, and treat "algorithms" as abstract computational procedures or as embedded in material objects (robots) or nonmaterial ones (recommendation systems). They classify four decision configurations (which evaluations are being compared) and show that the observed direction of preference can depend on which configuration a study measures; they explicitly call for hypotheses specific to the direction of preference and to moderators. Together, the two sources define the unresolved integrative question: when, why, and under which measurement conventions does the human-versus-AI preference reverse direction.

Problem Statement

Determine, with causal evidence, the specific conditions under which the direction of human preference between AI/algorithmic and human judgment reverses, and make the resulting account interpretable across studies. Concretely: (a) experimentally test whether the two dimensions of the capability–personalization framework of Qin et al. (2025) — perceived relative capability of AI and perceived necessity of personalization — causally control the sign of the preference effect, including whether manipulating them flips the observed direction (appreciation becoming aversion and vice versa) within the same decision context; (b) adjudicate the sources of the moderator heterogeneity they report — agent tangibility (tangible robots vs. intangible algorithms), outcome type (attitudinal vs. behavioral), experimental design (between- vs. within-subjects), and country-level education, internet penetration, and unemployment — deciding which are genuine causal boundary conditions and which are measurement or design artifacts; and (c) establish measurement configurations, in the sense of Jussupow et al. (2024), under which the observed direction of preference is comparable across studies. These three components jointly serve one overarching objective: a causally validated account specifying when people prefer AI over human judgment and when the reverse, such that a qualified researcher can predict the sign of the effect for a given decision context and measurement configuration. The completion boundary is that account together with its supporting causal evidence — a fully specified, experimentally testable causal account of when, and whether, the sign reverses under identified, validly manipulated conditions, together with an interpretable treatment of the tested moderators — not the enumeration of every possible decision context. The account specifies the manipulations of perceived relative capability and personalization necessity, the measures, and the predicted direction of preference under each manipulated condition.

References

  1. Qin, X., Zhou, X., Chen, C., Wu, D., Zhou, H., Dong, X., Cao, L., & Lu, J. G. (2025). AI aversion or appreciation? A capability–personalization framework and a meta-analytic review. Psychological Bulletin, 151(5), 580–599. https://doi.org/10.1037/bul0000477
  2. Jussupow, E., Benbasat, I., & Heinzl, A. (2024). An integrative perspective on algorithm aversion and appreciation in decision-making. MIS Quarterly, 48(4), 1575–1590. https://doi.org/10.25300/misq/2024/18512
  3. Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033
  4. Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005
  5. Longoni, C., Bonezzi, A., & Morewedge, C. K. (2019). Resistance to medical artificial intelligence. Journal of Consumer Research, 46(4), 629–650. https://doi.org/10.1093/jcr/ucz013
  6. Kaiser, M., Sunstein, C. R., & Reisch, L. A. (2025). On preferring people to algorithms. Journal of Risk and Uncertainty, 71(3), 245–272. https://doi.org/10.1007/s11166-025-09466-6
  7. Hu, Y., & Ding, Y. (2026). AI appreciation effect: How contagious diseases influence medical AI adoption. International Journal of Research in Marketing. https://doi.org/10.1016/j.ijresmar.2026.03.001
  8. Lu, Z., Guo, J., Yuan, T., Zhang, Y., Yang, J., Du, Y., Chen, M., Xie, M., Xian, L., Cao, H., & Zhang, K. (2026). A double-edged algorithm attitude: How appreciation and aversion shape students' AI learning anxiety in higher education. Behavioral Sciences, 16(6), 932. https://doi.org/10.3390/bs16060932