When Do Citizens Resist AI Governance? Experimental Evidence from the US

• In Event: From Crypto to AI: Mobilization On and Against New Platforms and Technologies (Virtual Session)

Abstract

In November 2020, Californians voted in a referendum on a proposal to replace the state’s cash bail for pretrial release with an algorithmic system. Despite evidence regarding the potential effectiveness of the algorithmic system in reducing crime among released defendants without imprisoning additional people, voters decisively rejected this proposition by more than ten percentage points. What explains this rejection?

Previous studies examining public attitudes toward the use of algorithmic risk assessments in bail decisions point to concerns about unfairness as a key explanation. Yet the question is whether this public opposition reflects a more general aversion to algorithms or whether it relates instead to the specific use of algorithmic decision-making systems (ADS) in criminal justice, where people are particularly sensitive to issues of unfairness and injustice. This question is particularly important given the increasing use of ADSs in various other policy contexts with large social implications, including policing, education, and social welfare assistance.

Because previous efforts to explore public attitudes toward ADSs have centered on specific policy areas, we lack both the theoretical and empirical perspectives needed to assess when and under what conditions, people are willing to accept the use of ADS in the public sector, and how they weigh issues of effectiveness against fairness when evaluating these uses.

This study evaluates when and under what conditions, people are willing to accept the use of ADS in the public sector, and how they weigh issues of effectiveness against fairness when evaluating these uses. By developing a theory to account for variations in preferences across decision-making contexts and subjecting it to empirical tests using novel data from two experiments embedded in a national representative survey or the US population. In the first experiment, respondents are asked to evaluate various decision-making scenarios where the identity of the decision-maker and the decision context are randomly manipulated. All scenarios presented in the survey are based on real-world initiatives and chosen according to distinguishing factors derived from the theory: (1) the object of the decision (i.e., whether the population directly affected by the decision consists of individuals or collectives), and (2) the broader objective of the decision (i.e., whether the decision is designed to sanction or to privilege).

Considering that most of the algorithms currently employed in the public sector are support tools, the second experiment assesses how the use of algorithmic support systems affects the perceived legitimacy of human decisions assisted by ADS. Respondents are presented with several additional scenarios that vary randomly across two dimensions: the algorithmic assessment (i.e., favorable or not favorable to the individual) and the outcome of the decision made by the human decision-maker (i.e., whether it is contrary or congruent with the algorithmic assessment).

Experimental evidence indicates that views on the use of algorithms differ substantially depending on the decision context in which it is deployed. The analysis indicates that while the deployment of AI in public policy may be acceptable in some cases, it could trigger substantial opposition in other, theoretically predictable contexts. Overall, this paper provides novel insights into the way ADSs can be used in public policy and the political implications of this growing phenomenon.

Author

• Shir Raviv, Tel Aviv University