Hiring algorithms are effectively unaccountable because opaque models, closed procurement, and absent per-candidate records leave rejected applicants without recourse; the author proposes a multi-level 'contestable accountability' framework and nine propositions to shift policy from explaining algorithms to ensuring decisions can be answered.

Main Finding

The paper argues that algorithmic opacity in automated hiring is not primarily a technical problem of model explainability but a governance problem: accountability fails because decision-level and procurement-level opacities, institutional incentives, and timing mismatches prevent affected applicants from obtaining meaningful redress. The author develops a multi-level framework of "contestable accountability" (four nested levels × three temporal moments), distinguishes model/system/decision opacity plus procurement opacity, and identifies two mechanisms—vertical and temporal displacement—that explain persistent governance failure. The policy implication is a shift from making algorithms explicable to making hiring decisions answerable.

Key Points

• Disaggregation of the “black box”:

  • Model opacity: internal function/feature weights of a trained model.
  • System opacity: socio-technical pipeline (pool construction, thresholds, rules, human integration).
  • Decision opacity: per-applicant inability to establish what happened in their case (was an automated assessment used, on what evidence, what score/result).
  • Procurement opacity (new): employers often contract for outputs and lack contractual access to models, training data, or validation—making the deploying organization itself an outsider.

• Opacity gradient and asymmetry:

  • Knowledge typically declines from vendor → employer → recruiter → candidate; candidates have least system knowledge but most at stake.
  • Vendors may sit outside employment duties but hold most system knowledge, generating accountability gaps.

• Why talent acquisition is distinctive (five features that complicate governance):

  • Decisions are one-shot and terminal.
  • Feedback is structurally absent (censored inference problem).
  • Subjects (applicants) are uninformed and unrepresented.
  • Harms are distributional/aggregate rather than clearly actionable in single cases.
  • Legitimacy is a market good; employers can prioritize appearing fair over being fair.

• Prevailing governance instruments reviewed and their limits:

  • Transparency disclosures: often satisfy formal requirements without creating actionable information for candidates, can backfire.
  • Explainability (post hoc explanations): addresses model opacity but not decision-level answerability or evidentiary needs.
  • Technical debiasing, human-in-the-loop review, algorithmic audits: each presumes institutional conditions (ongoing relationship, preserved records, access rights) that hiring lacks.
  • These instruments can be subject to “vertical displacement” (responding to redress demands with disclosure-oriented, low-cost fixes) and “temporal displacement” (governance effort concentrated pre-deployment while harms arise ex post).

• Multi-level contestable accountability framework:

  • Four nested levels: epistemic (knowledge), procedural (processes), institutional (roles & rules), contestatory (routes for challenge).
  • Three temporal moments: ex ante design, in itinere operation, ex post decision.
  • The framework yields nine testable propositions to guide empirical work and policy.

• Legal/regulatory landscape (as of Aug 2026):

  • European Union: shifting/highly volatile (EU postponed some high-risk obligations; CJEU decision in Dun & Bradstreet Austria expanded decision-level information duties).
  • United States: federal pullback from prior guidance; litigation (e.g., Mobley v. Workday) testing vendor liability and agent theories.
  • India: adopted a largely voluntary approach.

Data & Methods

• Type: Conceptual/theory-building paper (no new empirical dataset).
• Methods used:
  • Systematic literature synthesis across management, AI fairness, legal scholarship, and institutional theory.
  • Conceptual disaggregation of opacity types and institutional features of hiring.
  • Legal/regulatory survey up to August 2026 (cases and regulatory developments cited).
  • Development of a multi-level theoretical framework and derivation of nine falsifiable propositions for future empirical testing.
  • Use of illustrative cases and prior empirical findings (e.g., studies on disclosure compliance, ethnography of model development, Amazon reported case, litigation examples) to ground arguments.

Implications for AI Economics

• Market structure and vendor power:

  • Procurement opacity implies substantial information asymmetries between vendors and employers; vendors can internalize rents and externalize harms, affecting bargaining, competition, and market efficiency.
  • Vendor liability (if established in litigation) could shift incentives, altering pricing, contract terms, and investment in transparency/record-keeping.

• Labor-market matching and search frictions:

  • Automated exclusion of qualified candidates (even if unmeasured) increases frictions and may reduce effective match quality and overall labor market welfare.
  • Absence of feedback loops (censoring of counterfactual outcomes) impedes model calibration and market learning, potentially entrenching suboptimal screening equilibria.

• Signaling and reputation externalities:

  • Employers compete on perceived fairness; this reputational competition can substitute for substantive fairness, creating a signaling equilibrium where low-cost disclosure serves as a market signal (decoupling).
  • Aggregate reputation effects can mask distributional harms; economists should model reputation as a public-good externality that firms use strategically.

• Principal-agent and moral hazard:

  • Employers delegating screening to vendors face principal-agent problems; procurement opacity exacerbates moral hazard where vendors need not internalize downstream harms.
  • Contract design (access to validation, evidentiary record retention, audit rights) becomes a critical economic lever to align incentives.

• Regulatory economics and compliance costs:

  • Regulations focused on model explainability may misallocate compliance cost without improving candidate-level answerability; effective regulation may need to mandate per-candidate record retention, evidentiary access, and contractual transparency.
  • Timing of regulatory intervention matters—ex post contestability mechanisms (audits, complaint routes, evidentiary preservation) likely have higher social return than purely ex ante model documentation.

• Empirical agenda for AI economics:

  • Measure prevalence and consequences of procurement opacity across industries.
  • Quantify welfare loss from decision opacity (counterfactual inference, lost matches).
  • Estimate how vertical/temporal displacement change firm behavior and social outcomes.
  • Evaluate policy treatments: mandated evidentiary records, vendor disclosure requirements, liability rules—on market entry, costs, employment outcomes, and discrimination externalities.
  • Study how reputation-mediated incentives interact with formal regulation to produce decoupling equilibria.

Overall, the paper reframes the economic problem: reduce information asymmetries and align incentives not merely by opening models, but by making individual hiring decisions evidentially answerable and by reshaping procurement and institutional arrangements that currently block contestability.