Will an artificial intelligence really control your tax return?

Machine learning is one of the most widespread branches of artificial intelligence. The tax benefits are not omitted and the control of declaration errors (missing or erroneous) works well with this type of technology. And unlike this urban legend, AI is great for detecting undeclared swimming pools …

© stock.adobe.com

AI to track scammers

The tracking of omissions or errors in tax returns is intensified with these new tools. This software now uses technologies found in companies that need to process large amounts of information: machine learning. Thus, every year, the tax administration strengthens its use of artificial intelligence to combat fraud and improves the efficiency of its services. In 2021, the proportion of controls targeted by AI was 45% compared to 32.5% a year earlier. In the last few weeks, the teams in the Directorate-General for Public Finance (DGFIP) have familiarized themselves with a new tool. Its name: Galaxy. Approved on March 11 by the National Commission on Computing and Liberties (Cnil), the use of this new computer software for automated taxpayer data processing should allow tax authorities to better track down fraudsters.

Amounts received after tax audit

Years Amounts received in billions of euros N-1 variations in%
2021 10.7 billion euros +37.35%
2020 € 7.79 billion -29.18%
2019 € 11.0 billion reach

Galaxy: stars in the eyes of tax auditors

Its spread in DGFIP’s premises is part of the PILAT project, which was launched in 2018 to modernize the tax control system. The decree of March 11 describes this new weapon in the fight against fraud as “a visualization tool at national level on the one hand of the links between professional entities (links of participation) and between professional entities and natural persons (links of director, partner or shareholder), and on the other hand elements of context about these persons’ property and taxation situation“. Galaxy thus makes it possible to aggregate data such as the reference tax income, the identity of the spouse or even tax liabilities for individuals. But also the Siren number, the legal status, the category of turnover as well as the taxes and duties to which legal persons are subject.

45% of the controls are targeted at artificial intelligence

In 2021, € 13.4 billion in adjustments were notified to individuals and businesses, and € 10.7 billion was raised. A level close to the 2019 record (11 billion), which the tax administration largely attributes to “the use of data mining to better target tax audits”, while the share of audits targeted by artificial intelligence and data mining was almost 45% last year compared with 32.5% in 2020 and 21.95% in 2019. For 2022, it expects 50% of targeted controls thanks to artificial intelligence.


The tax authorities do not hide their intention “to increase the use of artificial intelligence”. And not just to fight fraud. Since March 2021, the deployment of digital assistants “by recovering fines has made it possible to automatically process the accounting posting of received bank transfers and to agents to improve the results of enforcement through the intensification of prosecution”, explains DGFIP.


Also in 2021, the IRS began using artificial intelligence to “help contact center agents respond to user emails about payment issues.” It has also implemented the “Weak Signals” project, which is based on an algorithm capable of “targeting companies’ weaknesses in order to establish support actions as early as possible”. This system made it possible to detect 23,227 companies in difficulty. Of these, 11,952 were selected to be contacted and offered support.

Swimming pool detection takes water

On social networks, exposing swimming pools or black garden sheds is more worrying than fraud in declaring cross-interests between a one-man business and charging housing allowance. That is to say, whether this kind of fraud should be important. But scammers can be temporarily reassured, AI does not allow in a sufficiently relevant way to automate the discovery of swimming pools or large-scale garden sheds. The scammers have anticipated and the use of tarpaulins in the same color as the environment close to the camouflaged property still works well.


Thus, in the experimental phase, DGFIP’s “Innovative Land” project aims to use artificial intelligence to automatically detect buildings and swimming pools that have not been declared from aerial photographs. Problem, the device tested in nine departments would not really be developed, with error rates of 30% at this stage. The tool will e.g. identify taxable swimming pools when they are above ground non-taxable swimming pools. Trade unions also report situations where artificial intelligence sees “buildings to be taxed instead of tarpaulins, roads, sidewalks, parking lots, slabs“, Stated to the Parisian Frédéric Scalbert, representative of the CGT.

Leave a Comment