Substance, Form, and Machines: AI Assistance and Human Normative Judgment in Taxation
TL;DR
The article proposes an accountable tax-AI architecture built on task classification, authoritative and time-stamped sources, explicit assumptions, traceable evidence, mandatory escalation, reason-giving, taxpayer contestability, and post-deployment audit that treats AI as an evidentiary and analytical assistant while preserving the human responsibility that makes tax law intelligible, challengeable, and legitimate.
Abstract
Artificial intelligence is becoming embedded in tax administration, professional advisory work, and digital compliance systems. Its expansion raises a central institutional question: when does a tax task permit reliable computational assistance, and when does it require human normative judgment? This article develops a normative-doctrinal account of that boundary. Drawing on legal theory, comparative anti-abuse doctrine, transfer-pricing guidance, empirical research on large language models, and contemporary AI-governance frameworks, it argues that AI can be valuable for structured retrieval, arithmetic, document synthesis, anomaly detection, and bounded classification. Yet a formally plausible output does not establish a legally justified tax outcome where law requires assessment of economic substance, business purpose, conduct, evidence, proportionality, fairness, or the reasoned exercise of public authority. The article proposes an accountable tax-AI architecture built on task classification, authoritative and time-stamped sources, explicit assumptions, traceable evidence, mandatory escalation, reason-giving, taxpayer contestability, and post-deployment audit. The resulting model is neither technological refusal nor blind automation. It treats AI as an evidentiary and analytical assistant while preserving the human responsibility that makes tax law intelligible, challengeable, and legitimate.
