State of use of AI tax systems
The Belgian Tax Administration (SPF Finances) has been using machine-learning algorithms since at least 2014, according to reports of the OECD and IOTA.
Incidental mentions attest that some models may have been used tested and used already earlier than 2014.
What functions are performed with AI?
The machine-learning algorithms of the Belgian Tax Administration perform a wide range of functions:
1. AI web-scraping: the SPF Finances makes use of machine-learning to automatically collect taxpayer data from e-commerce and e-sharing platforms, e.g. Amazon, Airbnb, eBay, 2emeMain, etc.
2. Social Network Analysis (SNA): the system is used to visually represents a network of individual taxpayers. The software represents a network of taxpayers as a combination of nodes (representing taxpayers or points of interests) and vertices which quantitatively and/or qualitatively measure relations between the nodes, using graph theory.
3. Transaction Network Analysis (TNA): the SPF Finances uses a machine-learning detection tool to flag in real-time and potentially block suspicious transactions for VAT purposes. The TNA system is inspired by an approach used since 2002 by the Special Tax Inspectorate. Since 2019 TNA has been put at the disposal of all EU Member States (see EU Report) by the Belgian SPF Finances as chair of the EU Fiscalis expert group.
4. Internal risk-management: The SPF Finances uses a suite of algorithms to predict the risk that taxpayers do not pay their taxes due, following a letter from the bailiff (‘Pegasus’), or following a call from the outbound call center (‘Iris’). These algorithms assist the SPF with their internal case management, and predict what course of action is most appropriate for the administration, based on historical taxpayer data, e.g. for taxpayers who are notoriously compliant/non-compliant these models prescribe a more coercive/cooperative course of action, and vice versa.
5. External risk-management (risk-scoring): The SPF Finances uses a suite of algorithms to predict specific risks of non-compliance of individual taxpayers, ‘Hermes’ predicts the risk of bankruptcy within a 12 months period for legal and self-employed persons, ‘Delphi’ predicts the solvency rate for natural, legal and self-employed persons. The SPF Finances also uses models to segment taxpayers into categories of risks to develop their annual audit plans, and select taxpayers with high-risks of non-compliance for further audits by human tax officials.
6. Nudging: the Belgian Tax Administration uses machine-learning to adapt the language of standard communication of taxpayers based on an analysis of individual taxpayer data.
What data can be processed by these systems?
The data used for these models is not specified. In accordance with the ‘Loi du 3 août 2012 portant dispositions relatives aux traitements de données à caractère personnel réalisés par le Service public fédéral Finances dans le cadre de ses missions‘ all taxpayer data is stored in a central data warehouse, and can be used for the development or use of algorithms.
Art. 327 of the Belgian Income Tax Code establishes that other governmental institutions must provide to tax officials any information that would be deemed relevant for the collection and levy of taxes. This information cannot be communicated or exchanged without the ‘express authorisation’ of the Federal Public Service. As provided by Prof. De Raedt, this express authorisation requirement was adapted by Art. 72 of the ‘Loi du 5 Septembre 2018 instituant le comité de sécurité de l’information’.
Are these systems regulated by specific norms?
There are no specific legal norms which regulate the use of machine-learning algorithms by the SPF Finances.
The use of AI by tax administrations is regulated as any other data processing by Articles 4 and 5 of the ‘Loi du 3 août 2012 portant dispositions relatives aux traitements de données à caractère personnel réalisés par le Service public fédéral Finances dans le cadre de ses missions’.
