Source-linked AI summary
How Do Fairness Definitions Fare? Examining Public Attitudes Towards Algorithmic Definitions of Fairness
Nripsuta Saxena, Karen Huang, Evan DeFilippis, Goran Radanovic, David Parkes, Yang Liu
TL;DR
Algorithmic fairness lacks an agreed definition, so this paper studies how ordinary people judge three alternatives in loan-allocation scenarios, with and without applicants’ race. Across two online experiments, calibrated fairness was generally preferred, and some race-sensitive judgments supported affirmative action.
Problem
Researchers have proposed multiple algorithmic fairness definitions but lack agreement about which is appropriate, while public views remain relatively understudied.
Method
Across two online loan-allocation experiments, the paper compares public judgments of three operationalized fairness definitions while varying repayment-rate similarity and applicants’ race.
Results
Participants broadly preferred the Ratio decision, indicative of calibrated fairness, and Study 2 found some support for affirmative action.
Takeaways & Limitations
Public attitudes can inform dialogue between technologists and ethicists when designing algorithms that make consequential decisions.
Takeaways & Limitations
The findings are an initial contribution to a broader research program, and future work should test other definitions, decision contexts, and reasons sensitive information changes judgments.
Abstract
from arXiv · showhide
What is the best way to define algorithmic fairness? While many definitions of fairness have been proposed in the computer science literature, there is no clear agreement over a particular definition. In this work, we investigate ordinary people's perceptions of three of these fairness definitions. Across two online experiments, we test which definitions people perceive to be the fairest in the context of loan decisions, and whether fairness perceptions change with the addition of sensitive information (i.e., race of the loan applicants). Overall, one definition (calibrated fairness) tends to be more preferred than the others, and the results also provide support for the principle of affirmative action.
Introduction
Algorithmic decisions increasingly affect high-impact domains, but disagreement persists over how fairness should be defined. Because the public is affected by these systems, the paper examines public attitudes toward competing fairness criteria.
- Algorithms increasingly shape decisions in loans, hiring, bail, and university admissions, creating broad societal implications.
- Researchers disagree about which algorithmic fairness definition is most appropriate, and some definitions cannot coexist.
- Relatively little research has examined how the general public evaluates fairness criteria in algorithmic decision-making.
- The paper tests public perceptions of different definitions to inform context-sensitive fairness research and dialogue between technologists and ethicists.
Definitions of Fairness
The paper studies distributive fairness in outcomes by comparing three computer-science definitions that differ in how individual characteristics should affect decisions.
- The study treats fairness as distributive justice concerning the fairness of outcomes.
- Fairness is framed as avoiding bias based on characteristics that are irrelevant in a particular decision context.
- The experiments examine task-specific similarity through loan repayment rates and sensitive attributes such as race.
- The paper operationalizes three fairness definitions as distinct loan-allocation choices understandable to ordinary participants.
- Calibrated fairness selects individuals in proportion to merit and, in this setting, is stronger than the other tested definitions.
Overview of Present Research
The research asks when people prefer one fairness definition over another, focusing on repayment-rate similarity and the influence of applicants’ race. Two online loan experiments test these perceptions in a divisible-resource setting.
- The central question is when people endorse one algorithmic fairness definition over another.
- The study tests how support for the three definitions changes as applicants become more or less similar in loan repayment rates.
- All three definitions treat race as irrelevant conditional on the task-specific metric, but affirmative-action reasoning may make race matter to distributive judgments.
- Across two online experiments, participants evaluate loan allocations, with Study 1 varying repayment-rate similarity and Study 2 adding applicants’ race.
- The fairness definitions were formalized for indivisible favorable decisions and therefore had to be interpreted for allocating divisible loan money.
Study 1 (No Sensitive Information)
Study 1 tests fairness judgments without race information by varying applicants’ repayment-rate similarity and comparing three allocation rules. Participants generally preferred the ratio allocation, while equal treatment was favored when repayment rates were nearly equal.
- Procedure: Study 1 presents participants with loan applicants whose repayment rates differ by 5%, 30%, 80%, or 80 percentage points across four treatments.
- Procedure: The study compares All A, Equal, and Ratio allocations, each designed to distinguish the three fairness definitions’ allowable outcomes.
- Procedure: The experiment measures perceptions of allowable outcomes rather than directly testing the fairness definitions themselves.
- Results: The Ratio decision was rated more fair than Equal in every treatment, supporting H1A.
- Results: The Ratio decision was rated more fair than All A in Treatments 1 and 2, providing partial support for H1B.
- Results: When repayment rates differed by 5%, Equal was rated more fair than All A, supporting H2 and indicating that applicants were viewed as similar enough for equal treatment.
- Results: All A was rated more fair than Equal in Treatment 3 but not Treatment 4, while no significant Ratio-versus-All A difference appeared in Treatments 3 and 4.
- Discussion: The Ratio decision aligned with calibrated fairness, whereas Equal was always aligned with treating similar people similarly.
Study 2 (With Sensitive Information)
Study 2 tested whether adding candidates’ race changes fairness judgments of three loan-allocation decisions. Participants generally preferred proportional allocation, while race altered some comparisons between equal and winner-take-all allocations.
- Procedure: Study 2 added candidates’ race to loan repayment rates and tested the same three allocation choices and hypotheses as Study 1.The study used the same experimental paradigm, with race randomized so either the white or black candidate had the higher repayment rate.
- Results: Participants consistently rated the “Ratio” decision as fairer than both “Equal” and “All A,” regardless of which candidate was white or black.The Ratio decision divided the $50,000 in proportion to the candidates’ loan repayment rates.
- Results: When repayment-rate differences were larger, participants rated “All A” as fairer than “Equal” only when the higher-repayment candidate was black.When the higher-repayment candidate was white, participants did not rate these decisions differently.
- Discussion: The authors interpret this race-contingent preference as support for more decisive allocations benefiting historically disadvantaged individuals with higher repayment rates.They describe this pattern as a boundary condition of H3 and connect it to affirmative-action reasoning.
Conclusion
The paper finds broad public preference for proportional loan allocation, supporting calibrated fairness over similarity-based and meritocratic definitions. It also finds some support for affirmative action and identifies several directions for extending research on public fairness judgments.
- Conclusion: People broadly preferred the “Ratio” decision, indicating support for calibrated fairness over treating similar people similarly and meritocratic definitions.The paper presents calibrated fairness as the strongest of the three definitions in this loan-allocation context.
- Conclusion: Study 2 also found some support for the principle of affirmative action in algorithmic fairness judgments.Race affected fairness attitudes in some loan-allocation treatments.
- Future research: Future research could test additional definitions, distinguish human from algorithmic decisions, examine other decision contexts, and study why sensitive information changes fairness perceptions.Suggested contexts include university admissions and bail decisions, alongside research on incorporating public views into algorithm design.
- Future research: The findings are presented as the start of a research program because ordinary people’s moral judgments can be inconsistent or sophisticated.Future work could examine whether moral-reasoning interventions influence fairness judgments.
Appendix
The appendix describes survey eligibility, study participation rules, and the organization of the appendix materials.
- MTurk workers had to be located in the United States to participate.
- Workers could participate in only one of the two studies.
- The appendix covers study questions, respondent demographics, and participant counts across four treatments.
Questions from the studies
The studies used separate questions, with Study 1’s question shown in Figure 4 and Study 2’s question shown in Figure 5.
- Study 1’s question is presented in Figure 4.
- Study 2’s question is presented in Figure 5.
Demographics questions
After answering their study question, MTurk workers were invited to answer voluntary demographic questions.
- Every worker answered the question for the study in which they participated.
- Most respondents answered some or all voluntary demographic questions, exceeding 90%.
1. What state do you live in?
The demographics questionnaire asked respondents to identify themselves, report education, describe their community, and provide related demographic information.
- Respondents were asked to identify themselves.
- The questionnaire asked about respondents’ highest completed level of schooling.
- Race response options included American-Indian or Alaskan Native, Asian, and Asian-American.
- Another race option was Native Hawaiian or other Pacific Islander.
- Respondents could select Other and specify an additional identity.
- Respondents were asked what type of community they lived in, including a city or urban community.
6. What is your age?
The paper reports participant demographic distributions for Studies 1 and 2, including age, education, gender, political affiliation, race, state, and residential breakdown.
- Study 1 participant demographics include age, education, gender, political affiliation, race, state, and residential breakdowns.
- Study 2 participant demographics include age, education, gender, political affiliation, race, and residential breakdowns.