AI tools can keep accountants productive while dulling judgment: a three-wave survey of accounting professionals finds that AI-related work pressure breeds both resilience that preserves efficiency and affective numbing that reduces ethical vigilance and professional skepticism, while greater AI autonomy and supervisor support are associated with healthier adaptation.

Main Finding

AI-induced work pressure in accounting produces two parallel adaptation pathways: (1) occupational resilience, which helps preserve sustained work efficiency, and (2) affective numbing, a state of reduced emotional responsiveness that does not reliably reduce efficiency but does erode ethical vigilance and professional skepticism. Organizational resources (AI autonomy and supervisor support) are directly associated with more resilience and less numbing, but they did not significantly moderate the pressure → numbing relationship.

Key Points

• New concept: "numb efficiency paradox" — efficiency metrics can remain stable while professionals become affectively detached, weakening the psychological conditions (sensitivity to anomalies, ethical discomfort, doubt) that underlie high-quality professional judgment.
• Dual adaptation pathways:
  • Occupational resilience: adaptive, associated with later sustained work efficiency.
  • Affective numbing: affective detachment under continued task functioning, associated with lower ethical vigilance and professional skepticism but not reliably with lower efficiency.
• Empirical results:
  • AI-mediated work pressure predicted both higher resilience and higher affective numbing (prospectively).
  • Occupational resilience predicted higher sustained work efficiency.
  • Affective numbing predicted lower ethical vigilance and lower professional skepticism; it did not significantly predict sustained efficiency after accounting for resilience and controls.
  • AI autonomy and supervisor support had beneficial main effects (higher resilience, lower numbing); hypothesized buffering (moderation) effects were not supported.
• Conceptual distinctions: affective numbing is distinguished from burnout, exhaustion, disengagement, automation bias, and moral disengagement — it specifically denotes blunted affective responsiveness despite maintained task completion.

Data & Methods

• Design: single three-wave longitudinal panel (no new recruits across waves).
  • T1 (baseline): n = 512 — measured AI-mediated work pressure, AI autonomy, supervisor support, baseline well-being, demographics/controls.
  • T2 (~6 weeks after T1): n = 417 — measured occupational resilience and affective numbing.
  • T3 (~12 weeks after T1): n = 361 — measured sustained work efficiency, ethical vigilance, professional skepticism.
• Sample: accounting and auditing professionals (U.S., U.K., Canada, Australia) recruited via Prolific; eligibility required active accounting/audit role, regular use of AI-enabled/automation tools, and accounting qualification or professional enrollment.
• Quality controls: attention checks, eligibility screening, consistency checks; attrition examined (retained participants had modestly lower AI pressure, higher supervisor support, and higher baseline well-being).
• Measurement & analysis: developed and validated measures for affective numbing and other constructs; prospective associations tested controlling for age, gender, tenure, weekly hours, and baseline well-being. Missing data handled via available-case estimation. Temporal separation (pressure → adaptation → outcomes) reduced same-source bias but does not establish causality.

Implications for AI Economics

• Productivity metrics can be misleading: Traditional economic measures of productivity (speed, throughput, completion rates) may understate the professional-quality risks introduced by AI adoption. Firms and regulators should incorporate measures of judgment quality (ethical vigilance, skepticism) when evaluating AI-enabled productivity gains.
• Hidden negative externalities: Affective numbing may increase risks (audit failures, misstatements, regulatory penalties, reputational harm) that are not captured by short-term efficiency gains. Economic evaluations of AI adoption should internalize these potential downstream costs.
• Organizational investments matter: Direct investments in worker-level resources (meaningful autonomy over AI use, supervisor support that legitimizes questioning AI outputs) are associated with more resilient adaptation and less numbing. Cost–benefit models of AI deployment should include the value of managerial and governance adjustments to preserve judgment quality.
• Labor market and human-capital effects: Persistent AI pressure may reshape skill demand — emphasis will increase on supervision, review, and judgment-preserving roles. Training, certification standards, and continuing professional education should address maintaining skepticism and ethical vigilance under AI-mediated workflows.
• Policy and regulation: Standard setters and regulators (e.g., auditing authorities) should consider guidance or requirements that go beyond tool accuracy and address human oversight, second-review rules, and organizational practices that prevent affective numbing. Incentive structures that reward pure throughput may need redesign.
• Research priorities for AI economics:
  • Link psychological states (numbing) to firm-level outcomes (audit quality, financial restatements, litigation, market reactions) to quantify economic costs.
  • Causal and field experiments testing interventions (autonomy redesign, supervisory training, mandated verification steps) and their costs/benefits.
  • Dynamic models of AI adoption that incorporate hidden quality risks, learning, monitoring costs, and long-run reputational externalities.
  • Cross-country and sectoral studies to estimate heterogeneity in numbing risk, institutional mitigation capacity, and economic impacts.

If you want, I can draft suggested metrics and practical checkpoints firms could use to monitor affective numbing (e.g., anomaly-review rates, second-opinion frequency, measures of ethical-vigilance in performance reviews) and estimate their likely costs relative to efficiency gains.