When identical advice is merely relabeled as coming from an algorithm rather than a person, laypeople weight it more heavily — the 'algorithm appreciation' effect of Logg, Minson and Moore (2019). What drives this effect, and when does it hold, remain unresolved. The source paper's future-research agenda asks: how people's lay 'theory of machine' — their perceptions of how algorithmic and human judgment differ in input, process, and output — shapes reliance on algorithmic advice; whether transparency of the algorithm's operations and people's more sophisticated understanding of them raises or lowers appreciation, given that complex operations become inscrutable and that, especially for the innumerate, such information might raise more concerns and doubts than it assuages; how beliefs about others' expertise affect advice uptake, whether human experts are preferred to algorithms, and how a lay theory of expertise compares with a lay theory of machine; and whether reliance on demonstrably high-quality aggregated algorithmic advice can be increased before any performance information is received. A substantial body of follow-up work has multiplied boundary conditions and partial reconciliations of algorithm aversion and appreciation, but the source's open directions remain unresolved.

Background

Since the 1950s, research on clinical versus statistical prediction has found that simple formal models often match or outperform expert human judgment; Dawes (1979) famously showed that even 'improper' linear models, built from roughly equal weights, can rival experts, partly because formulas are consistent and noise-free while human judges are not. This raises a behavioral question: do people actually use algorithmic advice? One literature answered negatively: Dietvorst, Simmons and Massey (2015) documented 'algorithm aversion' — people abandon an algorithm after seeing it err, even when it demonstrably outperforms them. The standard laboratory instrument for studying advice uptake is the judge–advisor system (JAS), in which a judge first states an initial estimate, receives an advisor's estimate, and then gives a final estimate. Reliance is measured by the weight of advice, WOA=∣Jf−J0∣/∣A−J0∣WOA = |J_f - J_0|/|A - J_0|, the fraction of the distance between the judge's initial estimate J0J_0 and the advisor's estimate AA that the final estimate JfJ_f travels toward the advice (WOA of 0 means ignoring the advice, 1 means adopting it fully). Logg, Minson and Moore (2019) used this paradigm across six experiments whose central manipulation was whether the advice's source label read 'an algorithm' or 'a person', with advice content held constant within each comparison. Contrary to the predictions of behavioral-science researchers themselves, lay participants gave reliably more weight to advice labeled as algorithmic (algorithm appreciation) when estimating visual stimuli and forecasting song popularity and romantic attraction; the preference persisted when participants chose directly between an algorithmic and a human advisor, waned when the alternative was the participant's own estimate, and was attenuated, though not reversed, among national-security forecasting professionals, who discounted advice from both sources. In a normative-benchmark study reported in their General Discussion, participants were told that the algorithmic advice was the aggregate of 314 estimates (normative WOA 1.0) while the human advice came from a single person randomly chosen from a pool of 314 (normative WOA 0.5); observed algorithmic WOA averaged 0.66, so participants still underweighted demonstrably high-quality aggregated advice relative to the normative prescription. The paper's own experiments varied the source label but never manipulated information about the algorithm's operations, transparency, or expertise, so they cannot say why the label matters; under the heading of a lay 'theory of machine' — people's perceptions of how algorithmic and human judgment differ in their input, process, and output — the authors explicitly identified as open problems the psychological mechanisms shaping reliance on algorithmic advice; how transparency of the algorithm's operations and people's more sophisticated understanding of them affect appreciation, including the possibility that, especially for the innumerate (numeracy = the ability to reason with numerical and probabilistic information), learning about the algorithm's operations might raise more concerns and doubts than it assuages, and that as operations become more complex they become inscrutable; how people's beliefs about others' expertise affect reliance on their advice, comparing a lay 'theory of expertise' with a 'theory of machine'; and how algorithmic advice can be presented to maximize its uptake prior to any performance information, including whether people expect algorithms and people to differ in how quickly they learn from mistakes. Subsequent reviews (Glikson & Woolley, 2020) list the tension between algorithm aversion and algorithm appreciation among the field's unresolved controversies, and later empirical work has shown the direction of the effect to be highly sensitive to framing (Hou & Jung, 2021), elicitation method (Logg & Schlund, 2024), task type (Castelo, Bos & Lehmann, 2019), task difficulty (Bogert, Schecter & Watson, 2021), and performance disclosure format (You, Yang & Li, 2022) — but the source's future-research directions remain open.

Problem Statement

The problem, as posed by the future-directions discussion of Logg, Minson & Moore (2019), is to replace the bare observation that people weight identical advice more when it is labeled 'algorithm' with a literature-grounded account of why and when they do so that, for each direction, states a falsifiable mechanism or boundary condition and the evidence that would confirm or refute it. The paper's future-research agenda (Section 6.3, 'Future research: Theory of machine') identifies these four open, scientifically linked directions: (1) Theory of machine — people's lay perceptions of how algorithmic and human judgment differ in their input, process, and output, as the theoretical framework for examining the psychological mechanisms that shape reliance on algorithmic advice. (2) Algorithmic process — how transparency of the algorithm's operations and people's more sophisticated understanding of them affect algorithm appreciation; the paper notes that as algorithmic operations become more complex they become inscrutable, and that especially for the innumerate, learning about the algorithm's operations might raise more concerns and doubts than it assuages, so disclosure may raise or lower reliance and may depend on the decision-maker's numeracy. (3) Expertise process — how people's beliefs about others' expertise affect reliance on their advice; whether people prefer human experts (as opposed to peers) over algorithms, and whether a lay 'theory of expertise' and a lay 'theory of machine' are the same or different constructs in governing advice uptake. (4) Room for increased reliance on algorithms — given that participants in the paper's normative-benchmark study still underweighted demonstrably high-quality aggregated algorithmic advice (observed mean weight of advice, WOA — the fraction of the distance between the judge's initial estimate and the advice that the final estimate travels, of 0.66 against a normative 1.0), how algorithmic advice can be presented in a way that maximizes its uptake prior to any performance information; and whether people expect algorithms and people to differ in how quickly they learn from mistakes. These four directions are the jointly mandatory, independently refutable components of the deliverable and together constitute its completion boundary: the account is resolved only when it addresses each direction, and resolving any one direction settles only that component while leaving the others open.

References

  1. Jennifer M. Logg, Julia A. Minson, Don A. Moore (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes 151:90–103. DOI: 10.1016/j.obhdp.2018.12.005. https://doi.org/10.1016/j.obhdp.2018.12.005
  2. Robyn M. Dawes (1979). The robust beauty of improper linear models in decision making. American Psychologist 34(7):571–582. DOI: 10.1037/0003-066X.34.7.571. https://doi.org/10.1037/0003-066X.34.7.571
  3. Berkeley J. Dietvorst, Joseph P. Simmons, Cade Massey (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General 144(1):114–126. DOI: 10.1037/xge0000033. https://doi.org/10.1037/xge0000033
  4. Ella Glikson, Anita Williams Woolley (2020). Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals 14(2):627–660. DOI: 10.5465/annals.2018.0057. https://doi.org/10.5465/annals.2018.0057
  5. Noah Castelo, Maarten W. Bos, Donald R. Lehmann (2019). Task-dependent algorithm aversion. Journal of Marketing Research 56(5):809–825. DOI: 10.1177/0022243719851788. https://doi.org/10.1177/0022243719851788
  6. Yoyo Tsung-Yu Hou, Malte F. Jung (2021). Who is the expert? Reconciling algorithm aversion and algorithm appreciation in AI-supported decision making. Proceedings of the ACM on Human-Computer Interaction 5(CSCW2):1–25. DOI: 10.1145/3479864. https://doi.org/10.1145/3479864
  7. Eric Bogert, Aaron Schecter, Richard T. Watson (2021). Humans rely more on algorithms than social influence as a task becomes more difficult. Scientific Reports 11:8028. DOI: 10.1038/s41598-021-87480-9. https://doi.org/10.1038/s41598-021-87480-9
  8. Sangseok You, Cathy Liu Yang, Xitong Li (2022). Algorithmic versus human advice: Does presenting prediction performance matter for algorithm appreciation? Journal of Management Information Systems 39(2):336–365. DOI: 10.1080/07421222.2022.2063553. https://doi.org/10.1080/07421222.2022.2063553
  9. Jennifer Logg, Rachel Schlund (2024). A simple explanation reconciles 'algorithm aversion' and 'algorithm appreciation': Hypotheticals vs. real judgments. SSRN working paper. DOI: 10.2139/ssrn.4687557. https://doi.org/10.2139/ssrn.4687557
  10. Andrej Gill, Robert M. Gillenkirch, Julia Ortner, Louis Velthuis (2024). Dynamics of reliance on algorithmic advice. Journal of Behavioral Decision Making 37(4):e2414. DOI: 10.1002/bdm.2414. https://doi.org/10.1002/bdm.2414
  11. Manhui Jin, Zhiyong Yang, Traci L. Freling, Narayanan Janakiraman (2025). The human superiority effect in advice taking: A multimethod exploration and implications for policy makers and governmental organizations. Journal of Public Policy & Marketing 44(3):350–369. DOI: 10.1177/07439156251320314. https://doi.org/10.1177/07439156251320314
  12. Peter Kotzian, Barbara E. Weißenberger, Sonja G. Prinz (2025). Not so averse after all: Behavioral mechanisms underlying the use of algorithms in managerial forecasting. International Journal of Human–Computer Interaction 42(7):4907–4931. DOI: 10.1080/10447318.2025.2543988. https://doi.org/10.1080/10447318.2025.2543988
  13. Yiwen Wang, Min Xu, Lin Zhao (2025). Exploring the cognitive mechanisms behind the adoption of algorithmic advice. Business & Information Systems Engineering 68(3):497–521. DOI: 10.1007/s12599-025-00925-7. https://doi.org/10.1007/s12599-025-00925-7
  14. Oliver Schilke, Martin Reimann (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes 188:104405. DOI: 10.1016/j.obhdp.2025.104405. https://doi.org/10.1016/j.obhdp.2025.104405
  15. Sander de Jong, Maarten W. Bos, Niels van Berkel, Maarten H. Lamers (2025). Algorithm appreciation or aversion: The effects of accuracy disclosure on users' reliance on algorithmic suggestions. Behaviour & Information Technology (online 1–20). DOI: 10.1080/0144929X.2025.2535732. https://doi.org/10.1080/0144929X.2025.2535732