Evidence Assessment in Tax Cases - a Foundation for Future AI-Based Decision Support

Background and Purpose

In tax proceedings, it is not only a matter of being right in substance but also of obtaining justice. The outcome of a tax case is largely determined by how evidence is assessed. An incorrect assessment of evidence can lead to additional taxation, tax surcharges and, in some cases, personal liability, with significant financial consequences for individuals and businesses. The individual is often at an informational and resource disadvantage vis-à-vis the public authorities, which makes the quality, traceability and predictability of the assessment of evidence a central issue of legal certainty.

Swedish law is based on the free submission and free assessment of evidence pursuant to Chapter 35, Section 1 of the Code of Judicial Procedure (1942:740, RB), while the requirements of the Administrative Procedure Act (2017:900) regarding legality, objectivity, impartiality, proportionality, documentation and reasoning, as well as the European Convention's requirement for a fair trial (Article 6), must be fulfilled. The balance between individualised assessment and systematic uniformity is particularly sensitive in tax cases.

Against this background, the project aims to develop a legal science foundation that makes the assessment of evidence in tax proceedings more methodologically sound and transparent, and enables the future development of tools based on generative AI technology (large language models (LLMs) such as ChatGPT and similar) that are legally secure and result in assessments that are as accurate as possible. The intention of this study is to lay a legal science foundation for the development of technical support for evidence assessment in tax proceedings.

State of Research and Knowledge Gap

Despite the central role of evidence assessment, there is limited systematic knowledge about how decision-makers reason and what constitutes a correct assessment of evidence in tax cases. There is generally a rather limited number of articles in Swedish tax journals dealing with this topic. The issue of evidence assessment and recent technology has not been addressed in the literature at all.

Internationally, the descriptive story model emphasises that decision-makers organise evidence into competing narratives and choose the most coherent one; in Swedish law, a holistic approach is expressed in RB 35:1, where the entirety of everything that has occurred must be examined conscientiously. At the same time, there is currently no method that combines the analytical capacity of large language models with the legal requirements that apply to evidence assessment in tax proceedings.

The above creates a clear research gap: before the technology can be used with legitimacy, a normatively and empirically grounded model for reliable evidence assessment in tax cases is required that can form the basis for future systems, in line with, among other things, the EU AI Act's requirements for transparency, risk management and human control.

Theory and Method

The project takes its theoretical starting point in descriptive general theories of evidence, particularly the story model, and combines these with legal dogmatic analysis of the burden of proof, standards of proof and the principles of evidence assessment in tax law, as well as administrative law requirements for objectivity, documentation, reasoning and party access.

Based on this foundation, an Evidence Assessment Framework (EAF) is developed that formalises decision nodes and reasoning structures for relevance, reliability, evidential strength, rebuttal, standard of proof and risk of error, and that explicitly supports traceability and controllability. The EAF is translated into a coding scheme that enables systematic analysis of court reasoning and can function as an annotation format for future datasets.

Methodologically, legal dogmatics is combined with empirical analysis of court reasoning. A selection of typical cases concerning evidence assessment issues is annotated independently by two assessors; inter-rater reliability is measured and calibrated; deviations are analysed to improve the framework. In parallel, a conceptual and technical preliminary study is conducted, resulting in a requirements specification for how a Swedish-language LLM-based decision support system with retrieval support (RAG) could be designed specifically for evidence assessment, including requirements for data governance, logging, explainability and human oversight.