State of use of AI tax systems
The Netherlands has, with Member States such as Belgium, Denmark or Sweden, always been at the forefront of automation and the use of technology to combat tax fraud and ensure tax compliance.
As early 2004, the Dutch Belastingdienst had developed an AI web-scraping tool ‘XENON’. In 2006, the Netherlands was one of the Member States piloting the EU FISCALIS network, enhancing cooperation and exchange of tax information between EU nations.
The Belastingdienst has always exhibited a strong incline toward technology, which may partially explain why two landmark cases of algorithmic misconduct, the toeslagenaffaire (See Publication) and SyRI (See Publication), occurred in the Netherlands.
What functions are performed with AI?
The Belastingdienst performs at least four functions with machine-learning:
- AI web-scraping: ‘XENON’: automatically collects data from HTLM scripts of websites, social media, e-commerce and e-sharing platforms to match it with data already present in the databases of the tax administration, through data matching. XENON comprises: a training tool, to calibrate the data to be scraped from HTLM scripts; a spider, which can follow any hyperlink in the script and scrape the data from connected sites; and, an identification tool, to identify website owners. The underlying objective of XENON is to detect unknown taxpayers and potential signals of non-compliance.
Sources indicate that Austria, Denmark, Sweden and other Member States have also used XENON to scrape taxpayer online data. - Social Network Analysis (SNA): the SNA system visually represents a network of taxpayers. SNA visually represents a network of taxpayers as a combination of nodes and vertices. Using graph theory, SNA is used to quantitatively and qualitatively measure relations between the nodes. The Belastingdienst uses SNA to detect group-level risks for various tax items, for instance it is particularly useful to detect missing-trader and carousel schemes. SNA was also used to detect ‘facilitators’, i.e. institutions which facilitated fraud, in the toeslagenaffaire.
- External risk-management (risk-scoring): the Belastingdienst uses a suite of risk-scoring algorithms, to segment taxpayers into risk categories and devise treatment strategies. An example of such risk-scoring system is Systeem Risico Indicatie or SyRI, a machine-learning system which predicts the risks of tax fraud or tax non-compliance associated with individual taxpayers subsequently selecting taxpayers above a certain threshold value for further audits. The use of this system was halted by the Court of the Hague in 2020.
- Nudging: the Belastingdienst makes uses of a nudging system which adapts the language used in standard communication to taxpayers, based on behavioural insights derived from scientific literature and inference from profiling of individual taxpayers. Reportedly, one of these nudging tools is targeting recent divorcees, as these taxpayers have been found to be a particular group prone to tax non-compliance.
What data can be processed by these systems?
The specific data used for the development and use of these tax machine-learning algorithms is not specified.
Art. 64 (1) (a-d) of the SUWI wet (now repealed) provides that any government records or data held by Dutch administrative bodies can be used for the purpose of preventing or combatting tax fraud.
Are these systems regulated by specific norms?
Among all the tax machine-learning algorithms aforementioned only SyRI was regulated by specific legislative norms, namely Art. 64 & 65 of the SUWI wet.
However, since Art. 65 of the SUWI wet was found to be in breach of Art. 8(2) of the ECHR, these algorithms are currently not regulated by specific legal norms.
