Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Can AI expose tax loopholes? Towards a new generation of legal policy assistants

Peter Fratrič, Nils Holzenberger, David Restrepo Amariles

TL;DR

The paper tackles the tax gap problem by introducing a hybrid neuro-symbolic system that combines a natural-language interface with a domain-specific language for planning to simulate multinational incorporation and tax-return decisions. It demonstrates how tax loopholes, exemplified by schemes like the Double Irish with the Dutch Sandwich, can be exposed through an agent-based, planning-driven exploration and inductive logic programming to inform policy design. The work provides theoretical guarantees and practical methods for improving social welfare via targeted restrictions, while highlighting limitations in NL rule formalization, computational scalability, and the need for human-driven policy decisions. Overall, the approach offers a pathway for publicly evaluating tax laws and supporting policy-makers with explainable, data-driven insights into potential loopholes and reforms.

Abstract

The legislative process is the backbone of a state built on solid institutions. Yet, due to the complexity of laws -- particularly tax law -- policies may lead to inequality and social tensions. In this study, we introduce a novel prototype system designed to address the issues of tax loopholes and tax avoidance. Our hybrid solution integrates a natural language interface with a domain-specific language tailored for planning. We demonstrate on a case study how tax loopholes and avoidance schemes can be exposed. We conclude that our prototype can help enhance social welfare by systematically identifying and addressing tax gaps stemming from loopholes.

Can AI expose tax loopholes? Towards a new generation of legal policy assistants

TL;DR

The paper tackles the tax gap problem by introducing a hybrid neuro-symbolic system that combines a natural-language interface with a domain-specific language for planning to simulate multinational incorporation and tax-return decisions. It demonstrates how tax loopholes, exemplified by schemes like the Double Irish with the Dutch Sandwich, can be exposed through an agent-based, planning-driven exploration and inductive logic programming to inform policy design. The work provides theoretical guarantees and practical methods for improving social welfare via targeted restrictions, while highlighting limitations in NL rule formalization, computational scalability, and the need for human-driven policy decisions. Overall, the approach offers a pathway for publicly evaluating tax laws and supporting policy-makers with explainable, data-driven insights into potential loopholes and reforms.

Abstract

The legislative process is the backbone of a state built on solid institutions. Yet, due to the complexity of laws -- particularly tax law -- policies may lead to inequality and social tensions. In this study, we introduce a novel prototype system designed to address the issues of tax loopholes and tax avoidance. Our hybrid solution integrates a natural language interface with a domain-specific language tailored for planning. We demonstrate on a case study how tax loopholes and avoidance schemes can be exposed. We conclude that our prototype can help enhance social welfare by systematically identifying and addressing tax gaps stemming from loopholes.

Paper Structure

This paper contains 27 sections, 2 theorems, 3 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Outline and contributions
  4. Building a model: Ambiguity as a feature
  5. Model and initial state
  6. States
  7. Actions and transitions
  8. Economic context
  9. Licensing contracts
  10. Commercial revenue
  11. Filing a tax return
  12. Utility function
  13. Revealing tax loopholes
  14. Search space exploration
  15. Legal summary statistics
  16. ...and 12 more sections

Key Result

Theorem 1

Let the utility $u$ be defined as in equation (eq:util), and let ${\phi(\gamma\langle s_0,s \rangle)}$ be proportional to the trajectory length ${||\gamma||}$. If ${p(\gamma\langle s_0,s \rangle)}$ is path-independent and ${||\gamma||}$ is minimal, then ${\gamma\langle s_0,s \rangle}$ is canonical.

Figures (6)

  • Figure 1: A subset of laws from the legal corpus are translated in a domain specific language (DSL), to define states, legal actions, and tax reductions in a socio-economic model. The user can define the initial state or interpret a new law from the corpus via a natural language interface constrained to produce DSL-compatible expressions. The exploration engine samples tax plans, conceptualized in the explanation module, and passed to the user.
  • Figure 2: Illustration of state initialization by the user. Probability of obtaining the correct response can be found in Figure \ref{['fig:llmStats']} for various language models.
  • Figure 3: Simulation results of LLM sampling for the initial state prompt formalization (left bar) and deductible prompt formalization (right bar). Temperature value of the sampler was selected by grid search for values $[0.01, 0.02, 0.05, 0.1 , 0.2 , 0.5 , 1.0 ]$. For other values not included in the plot, no correct answer was found.
  • Figure 4: Illustration of introducing a new tax reduction rule by the user. The probability of obtaining the correct response can be found in Figure \ref{['fig:llmStats']} for various language models.
  • Figure 5: Segmentation of the utility profile based on the slope value. To obtain the segments, numerical differentiation of the curve was performed and peak detection algorithm with values peak size of $3.85$ and minimal distance between peaks equal to $100$ was used to identify points with high negative slope.
  • ...and 1 more figures

Theorems & Definitions (7)

  • Example 1
  • Definition 1
  • Theorem 1
  • proof
  • Definition 2
  • Theorem 2
  • proof