State of use of AI tax systems

The Kingdom of Denmark has always exhibited a high level of maturity in terms of automation and digitalization of society, including the use of machine-learning. Prof. Brigitte Alfter on behalf of AlgorithmWatch reports that the Danish tax administration experimented with ‘traditional’ algorithms to automatically collect taxes (so-called EFI system – Et Fælles Inddrivelsessystem or One Shared Tax collection system) as early as 2005. The project was shelved in 2015 after major flaws with these systems were revealed. The State pension system, SLP-S, also used algorithms as early as 2001.

Regarding tax machine-learning algorithms, there are at least two publicly known use cases:

  • Firstly, Udbetaling Denmark – the centralised body in charge of welfare benefits payments such as unemployment benefits and child support, makes use of machine-learning to automate controls regarding controls and suspicions of fraud. The AI algorithms used by Udbetaling operates algorithmic-based selection of citizens suspected of fraud based on undeclared risk-factors. Birgitte Arent Eiriksson, deputy director of the legal think tank Jusititia, reports that Udbetaling has far-reaching competences in terms of data use to perform its prerogatives and that the institution seem to place efficiency as a higher priority over the respect for fundamental rights. 
  • The second known use case of machine-learning algorithms concerns the automatic valuation of real estate by the Danish tax administration. Reportedly, the model uses 19 variables, including proximity to facilities such as schools, parks, but also the level of pollution in the area.

What functions are performed with AI?

The above mentioned machine-learning algorithms perform at least two functions:

1. External risk-management (risk-scoring): the algorithm segments taxpayers into categories of risks and selects taxpayers with predicted high risks of fraud or non-compliance to devise an audit plan for the tax administration.

2. Automated calculation/verification of tax information (automated decision-making): the algorithm automatically calculates and/or verify pricing of real estate without any individual manual inputs from tax officials.

Another function cited in scientific literature is nudging, which adapts the language of standard communication sent to taxpayers, based on data analytics and predicted risk profiles, to incentivize these taxpayers to submit their tax returns within the allotted deadlines. Yet, so far, it seems that nudging remains at an experimental stage.

What data can be processed by these systems?

Regarding the machine-learning algorithm of Udbetaling, the data which can be leveraged for the development and use of the model is not specified. Nonetheless, as reported by Eiriksson, Udbetaling has access to a far-reaching range of data, as the ‘lov om Udbetaling Danmark’ (Law on Udbetaling Denmark) foresees the possibility to use any data of Udbetaling, but also the possibility of linking and registry merger, i.e. using data from other administrative bodies in Denmark (see Article 11, 12 and following).

Regarding the second algorithm mentioned, the model is based on 19 variables, including but not limited to, the sales prices of the real estate, its location, proximity to facilities, pollution in the area, etc.

Are these systems regulated by specific norms?

The algorithm of Udbetaling is regulated by the Lov om Udbetaling Danmark. However, as such the provisions of this law do not specifically regulate the use of the machine-learning algorithm but rather detail the use of data by the administrative body. Seemingly, the algorithm used by Udbetaling have been implemented by managerial decisions of Udbetaling.

The real estate valuation algorithm are regulated by the Ejendomsvurderingsloven(‘Property Valuation Act’).


Legislation :

Doctrine :