Individual Submission Summary
Share...

Direct link:

Poster #96 - When Transparency Works: Disclosure Timing and Public Responses to Algorithmic Decisions

Saturday, November 7, 12:00 to 1:30pm, Property: Boston Marriott Copley Place, Floor: 4th Floor, Room: Grand Ballroom - Salon EFG

Abstract

Governments are rapidly automating high-stakes administrative decisions in welfare, healthcare, and tax systems. To address public concerns, jurisdictions worldwide, including the EU, the United States, and China, have adopted governance frameworks centered on algorithmic transparency and human review of appeals. Yet evidence on these safeguards remains mixed. Some studies find that transparency increases trust, while others show that disclosure generates resistance. Why do procedural safeguards produce different responses across settings? We examine one overlooked explanation, disclosure timing. We test whether informing citizens of the involvement of automated decision-making systems (ADS) before a decision is made, compared with disclosure only after the decision has been made, changes citizen evaluations alongside algorithmic transparency and human review.

The study also addresses a broader question in policy design. Research on policy instruments has examined layering, sequencing, and combinations across policy episodes, but less attention has been paid to whether temporal parameters within a single policy episode shape instrument effectiveness. Fairness heuristic theory suggests they may. Procedural information available before an outcome is known provides a frame for interpreting the decision. Pre-decision notification establishes awareness of ADS involvement before the outcome, whereas post-decision disclosure introduces that information after an initial judgment has formed. This sequence may shape evaluations of procedural justice and institutional legitimacy, with effects distinct from those of transparency and human review.

We test these expectations using a preregistered 2 × 2 × 2 between-subjects experiment in the context of an automated review of eligibility for social assistance. Participants are randomly assigned to conditions varying disclosure timing (pre-decision notification versus post-decision disclosure), algorithmic transparency (detailed versus minimal explanation), and human review (staff-handled versus automated appeal). Procedural justice is measured through two validated dimensions, perceived decision quality and treatment quality. We also assess perceived legitimacy, immediate cognitive reactions, and changes in acceptance of automated decision-making.

Three findings emerge. First, pre-decision notification significantly improves perceived decision quality, treatment quality, and perceived legitimacy. It also directs immediate attention toward the substance of the decision and reduces the decline in acceptance of automated decision-making following an unfavorable outcome. These benefits weaken among respondents with prior unfair experience with ADS. Second, algorithmic transparency improves perceived decision quality and treatment quality, with a smaller improvement in perceived legitimacy. Its effects concentrate on citizens’ evaluations of the decision and its procedure. Third, human review of appeals does not significantly improve perceived decision quality or treatment quality, though it modestly increases perceived legitimacy. This pattern is consistent with an accountability signal operating at the institutional level.

These findings identify disclosure timing as a distinct procedural safeguard and offer a temporal perspective on mixed evidence in algorithmic governance. Notification, transparency, and human review address different aspects of citizen judgment, making their position within an administrative encounter relevant to policy design. For practitioners, the results support pre-decision notification of ADS involvement alongside understandable explanations and accessible human review. Incorporating notification into existing procedures provides an administratively feasible means of improving citizen evaluations of automated decisions.

Authors