Algorithmic and Artificial Intelligence Judgment Aids
Prevailing wisdom suggests that people tend to be averse to relying on algorithmically automated judgments in lieu of their own, a phenomenon frequently referred to as algorithm aversion. However, this result typically applies to situations where judges must decide between their own autonomous judgments, or choosing to allow an algorithmic agent, such as an AI model, to make a judgment on their behalf. On the other hand, there is evidence that when people are given the option to incorporate advice from algorithmic or human sources into their own judgments, they sometimes prefer algorithms. Even in cases where claim to have strong preferences between human and algorithmic advisors, that preference does not seem to impact how they use the two information sources in practice. Given the increasing ubiquity and accessibility of artificial intelligence (AI) and interactive large language models like ChatGPT, understanding how, when, and why humans interact with algorithmic advisors is of increasing importance. Given these results, we hypothesize that the primary hurdle in human adoption of AI advice is the initial exposure stage, but that if and when people “hear out” AI advisors, they will typically make good, if not optimal use of that information. We are currently designing studies to test this hypothesis, and exploring applications in domains such as smart cities, in which technology and AI are seamlessly integrated into an interconnected community.
Decline, Adopt or Compromise: Computational Models of Advice Taking
The dominant experimental paradigm in the advice utilization literature is known as the Judge Advisor System (JAS). In JAS experiments, participants are tasked with forming a quantitative judgment, such as estimating the age of a person in a photograph, or the probability each of several candidates will win an upcoming election. They are then given advice from an independent source, and asked to make a second judgment. Advice utilization is measured as a function of the difference between their first prior (to advice) and posterior judgments. A clear pattern emerges across studies that utilize this design: people appear to initially make a categorical decision, in which they choose between lower-effort options of declining or adopting advice, and a higher effort option that entails a compromise between the two. If they do choose to compromise, they must then take the extra step of judging how much to weigh their prior belief against the new advice. While this general pattern is clear, there is also evidence that the usage of the different strategies is heavily dependent both on individual tendency and on the context of the advice taking study. In many cases, decline decisions are prevalent, while in others, people are much more likely to adopt, and across all domains, reluctance to compromise can lead to suboptimal posterior judgments. We are exploring when, how, and why people choose each of these different strategies, and ways of intervening to encourage people to compromise in cases where it is most likely to benefit their posterior estimates.
Persuasion Without Prior Belief
A second consequence of the JAS paradigm is the causal effect of prior belief elicitation. Research clearly indicates that when prior beliefs are not elicited, the distribution of posterior estimates can be drastically different. This raises a problem: how can we study belief change in an ecologically valid way if measurement of prior belief alters the process? To accomplish this, we employ cutting edge Bayesian missing data imputation techniques. Results of two initial studies indicate the absence of prior belief measurement has dramatically different consequences in different contexts. In a continuous estimation task, in which participants estimated how many calories were in different foods, though people rarely declined advice, the decline, adopt, or compromise model still provided an adequate representation of the belief revision process. However, in a probability prediction task involving sports predictions, the model was a poor representation of the data. We are currently exploring new models to better represent the belief revision process for probability judgments in cases where people don’t explicitly report their prior beliefs.
Relevant Publications
Himmelstein, M. & Budescu, D.V. (2023).Preference for human or algorithmic forecasting advice does not predict if and how it is used. Journal of Behavioral Decision Making. 36(1), e2285.https://doi.org/10.1002/bdm.2285
Himmelstein, M. (2023). A Bayesian approach for using planned missingness to address measurement reactivity [Abstract]. Multivariate Behavioral Research. https://doi.org/10.1080/00273171.2022.2160296
Himmelstein, M. (2022). Decline, adopt or compromise? A dual hurdle model for advice utilization. Journal of Mathematical Psychology. 110, 102695. https://doi.org/10.1016/j.jmp.2022.102695
Himmelstein, M., & Budescu, D.V. (2022). Modeling boundary inflation in advice utilization for multinomial probability judgment [Abstract]. Multivariate Behavioral Research. https://doi.org/10.1080/00273171.2021.2009329
Morstatter, F., Galstyan, A., Satyukov, G., Benjamin, D., Abeliuk, A., Mirtaheri, M., Hossain, K. S. M. T., Szekely, P., Ferrara, E., Matsui, A., Steyvers, M., Bennet, S., Budescu, D., Himmelstein, M., … Abbas, A. (2019). SAGE: A Hybrid Geopolitical Event Forecasting System. Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, 6557–6559. https://www.ijcai.org/proceedings/2019/0955.pdf