Displaced Complementarity: Technology and Human Capacity in the Organizational Production of Government Transparency

• In Event: AI, Algorithms, and Decision-Making in Government

Abstract

Many public administrative processes unfold as sequences of distinct tasks, some amenable to automation, others dependent on human judgment. When governments automate one stage, evaluations typically focus on whether the technology helps or hinders where it is deployed. Yet technology that reshapes one stage also reshapes the demands placed on the stages that follow, and those downstream consequences have received little attention. Existing evidence is mixed, with some studies finding efficiency gains, others reporting errors, bias, or degraded quality. What is missing is not more evidence but a theory of when and why effects diverge.

We ask how the effects of automation vary across the stages of a sequential administrative process, and how human capacity conditions those effects. We develop a theory of displaced complementarity, which holds that automation at a routine upstream stage redistributes rather than reduces the demand for human judgment downstream. Because complementary activities are interdependent, a partial move that adopts technology without matching downstream investment in human capacity can push overall performance below the pre-adoption baseline rather than merely decrease its gains.

We test this theory in federal Freedom of Information Act request processing. This pipeline has an automatable upstream stage (record search and de-duplication) and a judgment-intensive downstream stage (review and respond), and institutional constraints systematically prevent staffing adjustments. Our agency-year panel covers 29 high-volume U.S. federal agencies from 2008 to 2024, drawing on Annual FOIA Reports and agency adoption records for de-duplication technology.

We estimate causal effects using a stacked difference-in-differences design that exploits staggered adoption between 2011 and 2024, with agency workload (request per staff) as the moderating variable. We report average treatment effects on the treated and distributed-lag specifications that trace dynamics over six years post-adoption.

Simple-track processing time, our upstream efficiency outcome, falls by roughly half from 21 to 10 working days. This deepens over three years and arrives more quickly in agencies with heavier workloads. Complex-track processing time, which spans both stages, shows a null average treatment effect. Yet its distributed lags are jointly significant, with signs that reverse across time horizons, consistent with upstream acceleration offset by a growing downstream review burden. Appeal rate, our downstream quality outcome, is unchanged on average after adoption. But in agencies with heavier workloads, requesters begin contesting decisions more often three years after adoption, and this keeps rising through Year 5 and beyond. The pattern signals a delayed erosion of decision quality that only surfaces once reviewers have carried the redistributed judgment burden long enough.

These findings reframe automation evaluation for public organizations. Metrics aligned with automation’s benefits respond on a faster timeline than those needed to detect its costs, producing optimistic early assessments that deepen organizational commitment before costs can be observed. Evaluations confined to the automated stage will systematically overstate benefits and miss downstream costs. Complementary staffing investment, not more technology, is the binding constraint on the organizational production of government transparency.

Author

• Hyorim Seo, Rutgers University - Newark
  Presenting Author