Belief in powerful AI tax-monitoring does not automatically deter corporate tax evasion; Indonesian firms who perceive stronger AI capability report higher intentions to engage in evasive tax planning, suggesting strategic adaptation rather than simple compliance.

Main Finding

Perceived AI capability (tax authorities’ AI competence as perceived by corporate tax decision-makers) is positively and significantly associated with corporate tax evasion intention in Indonesia. Contrary to the authors’ hypothesis, higher perceived AI capability correlates with higher, not lower, intent to evade taxes (t = 4.551, p < 0.001). The effect is substantive (f2 reported as meaningful), the model shows predictive relevance (Q2 > 0) and acceptable fit (SRMR < 0.08).

Key Points

• Research question: Does perceived AI capability reduce tax evasion intention among corporate taxpayers?
• Sample: 278 valid responses from corporate tax decision-makers in Indonesia.
• Main empirical result: Perceived AI capability → positive effect on tax evasion intention (hypothesized negative effect not supported).
• Interpretation offered by authors:
  • Perceptions of sophisticated AI surveillance appear to trigger more complex, strategic risk assessments by firms rather than uniformly increasing compliance.
  • When taxpayers view AI as capable, they may adapt tax planning strategies (gameability, selective concealment, exploitation of loopholes).
• Policy recommendation from authors: AI deployment in tax enforcement should be paired with transparency, legal certainty, and clearer risk communication to reduce strategic adaptation.
• Novelty: Integrates a technology-based psychological construct (perceived AI capability) into tax-behavior/intention literature.

Data & Methods

• Design: Quantitative, explanatory, cross-sectional survey.
• Respondents: Corporate tax decision-makers (n = 278).
• Measurement: Questionnaire items for perceived AI capability and tax evasion intention; standard validity/reliability checks reported (Cronbach’s α, Composite Reliability > 0.70; AVE > 0.50; Fornell–Larcker and HTMT confirm discriminant validity).
• Analysis: PLS-SEM using SmartPLS; two-stage evaluation (measurement and structural models); bootstrapping for significance testing.
• Key reported statistics:
  • Path coefficient for perceived AI capability → tax evasion intention: statistically significant, t = 4.551, p < 0.001.
  • Effect size: f2 indicates substantive contribution.
  • Predictive relevance: Q2 > 0.
  • Model fit: SRMR < 0.08.
• Optional analyses: Multi-group analysis by industry sector mentioned but no detailed subgroup results provided in the excerpt.

Implications for AI Economics

• Enforcement technology can produce unintended strategic responses: Economists modeling enforcement should account not only for changes in detection probability but also for how perceived algorithmic competence alters agents’ strategic behavior and investment in evasion tactics.
• Beliefs matter as much as objective capability: Perceptions of AI (shaped by partial information, media, or policy signals) can change incentives; policy effectiveness depends on managing beliefs (transparency, credible signaling).
• Design and deployment considerations:
  • Explainability and transparency may reduce “gameability” by clarifying detection targets and reducing informational asymmetries.
  • Randomized or less-predictable enforcement and complementary human audits may limit taxpayers’ ability to systematically adapt.
• Fiscal forecasting and welfare analysis: Models that ignore behavioral adaptation to perceived AI capacity may overestimate revenue gains from AI investments; dynamic, equilibrium models of taxpayer–authority interaction are needed.
• Policy toolbox: Pair AI systems with legal certainty, clear risk communication, penalty structures, and anti-avoidance rules—purely technical upgrades are insufficient and may shift evasion into more sophisticated forms.
• Research implications for AI economics:
  • Incorporate perceived-AI constructs into theoretical models of compliance and enforcement.
  • Evaluate whether different AI features (accuracy, explainability, scope) produce distinct behavioral responses.
  • Use experimental and longitudinal methods to identify causal channels (e.g., whether perceived competence raises confidence in evasion strategies, or reallocates evasion across instruments).

Limitations to keep in mind: cross-sectional self-report data (possible social desirability/common-method bias), single-country (Indonesia) context with nascent tax-AI implementation, and no behavioral outcome (intention rather than observed evasion). Future work should test behavioral outcomes, use experimental manipulation of perceived AI capability, and model dynamic taxpayer–authority interactions.