Tax Encounters in the Dark: Tax Law Meets AI’s Black Box

While artificial intelligence (AI) has been widely embraced in recent years, some commentators have raised concerns about the opacity of certain types of automated decision-making systems. It’s not always possible to know how a system arrived at a given decision. And that has important implications in many fields where visibility is a virtue, including tax.

Aleksandra Bal, senior product manager for VAT reporting and compliance solutions at Vertex, looks at how tax law is adapting to these new realities in a Tax Notes International article: Ruled by Algorithms: The Use of ‘Black Box’ Models in Tax Law. The article offers some compelling insights for tax executives who need to understand how tax is evolving in the AI era. In this article, Aleksandra explains:

Which AI models tend to limit visibility.

Alexandra helpfully categorizes the various machine learning models and their outputs. Among predictive models, neural networks and ensembles tend to be more complex and black-box in nature, she notes. “They generally deliver more accurate predictions, but it is difficult to understand why and how they produced a particular result, and the outcomes they generate are not intuitive.”

How Poland is mobilizing AI to fight VAT fraud.

In 2017, this EU member state initiated a comprehensive plan to close a VAT gap estimated at USD 11.2 billion for 2012. The initiative includes the use of secret algorithms to calculate a risk indicator for entrepreneurs, based on factors such as customer residence and the complexity of the entrepreneur’s ownership structure. 

How data protection regulations impact black box models.

Aleksandra focuses on the EU’s General Data Protection Regulation (GDPR), which directly addresses algorithmic decision-making. For example, three articles in GDPR establish a ‘right to explanation’ for EU citizens by requiring organizations handling the personal data to explain how an automated decision was reached.

How human rights legislation impacts the black box.

All EU member states are signatories to the European Convention for the Protection of Human Rights and Fundamental Freedoms (ECHR). The right to a fair trial, as guaranteed by ECHR, means that tax administrations must give reasons for their decisions and allow affected individuals access to their case file. “A decision made solely based on a black box model will likely conflict with those fundamental rights,” Aleksandra notes. The legal environment is complex, however, with multiple countervailing principles.

Indeed, complexity seems to be an inevitable outcome of tax law’s initial encounters with AI’s black box. But Aleksandra’s overview cuts through the complications, and she ends with some common-sense proposals on how to make AI work with an acceptable level of transparency.