State of use of AI tax systems
The tax administration in Sweden have been using machine-learning algorithms since at least as early as 2011, according to sources from the Forum on Tax Administration of the OECD.
What functions are performed with AI?
The Swedish tax administrations makes use of machine-learning algorithms to perform four functions:
- Risk detection: An unsupervised learning model is used to detect under-reported income. It uses cluster analyses and the nearest neighbours models to do so. In practice, the model compares the incomes reported by taxpayers to other taxpayers in similar situations to detect ‘outliers’, i.e. reported incomes which are abnormal when compared to the reported incomes of their nearest neighbours, hence taxpayers with similar characteristics.
- Web-scraping ‘EC-YES’: The Swedish Tax Administration uses a supervised learning model that on a regular basis automatically collects data from online webpages, commercial platforms, gambling websites and on the deepweb. It can also target specific websites on-demand of the authorities.
- Social Network Analysis (SNA): the SNA algorithm visually represents a network of individual taxpayers using graph theory, it represents a network of taxpayers as a combination of nodes for individuals or points of interests and lines which quantitatively and qualitatively measure relations between the nodes. Reportedly, the SNA model of the Swedish Tax Authorities was developed by the IT service provider SAS.
- Taxpayer assistance: The Swedish tax administration has developed a chatbot ‘Skatti’ that can answer taxpayers queries on personal taxation and population registration. According to Göran Sundin, digital strategist at the Swedish tax administration, Skatti handles around 15,000 chats per month and has handled more than 200.000 taxpayer queries since launch.
What data can be processed by these systems?
The data which has been used for the development of these machine-learning models, or the data continuously fed to these models, has not been specified by the Swedish tax administration.
Are these systems regulated by specific norms?
The models are not regulated by specific legal norms. So far these two models have been simply implemented on the basis of managerial decisions of the tax administration (cfr. Publications – eKasa case).
The Swedish government has recognised, in its first national AI strategy report released on May 2018, the need to develop additional legal norms and AI ethical principles to regulate the use of machine-learning algorithms. Yet, so far, such norms and principles have not been adopted in Sweden.
- OECD, Advanced Analytics for Better Tax Administration (2016), p. 21;
- CIAT, ICT as a strategic tool to leapfrog the efficiency of Tax Administrations (1st ed., 2007);
- Helene de Faire, ‘Skatti – Our Digital co-worker: Experiences from developing a chatbot using AI’ in IOTA, Applying New Technologies and Digital Solutions in Tax Compliance (IOTA, 2019), pp. 46-47;
- Government Offices of Sweden, National Approach to Artificial Intelligence (2018), p. 10 – available at: https://knowledge4policy.ec.europa.eu/ai-watch/sweden-ai-strategy-report_en - last accessed July 2022.
- Swedish National Center for Applied Artificial Intelligence (April 2019):https://www.ai.se/en/news/artificial-intelligence-improves-swedish-tax-agencys-customer-service-0 - last accessed July 2022.