Without the shadow of a visit, a ring of the doorbell or even a prying glance above their fence, some 140,000 owners of private inground swimming pools have received a letter since 2022 from the tax administration asking them to rectify their situation by declaring this chlorinated oasis. A question of respect for the local urban plan and the calculation of the property tax of which some are perfectly aware, but which others ignore, before being caught by the patrol. In this case, a very discreet controller assisted, as the State likes to point out, by artificial intelligence (AI).
Swimming pools are the much-publicized totem of the “Innovative Land” project, launched by the General Directorate of Public Finances (DGFiP) two years ago. Its objective: to identify these undeclared buildings quickly and automatically using image recognition operated by AI, on maps provided by the National Institute of Geographic and Forestry Information (IGN). The findings are then cross-referenced with declarations made to town planning services and other tax data. Until now, swimming pools have made excellent technological guinea pigs for the administration. On the one hand, this pretty azure blue denotes an aerial photo like stars on a black sky. And their geometry, often rectangles, is very recurring and relatively easy to detect. Finally, no one will pity the owners of personal pools, a symbol of a certain luxury of life, even if the error is in good faith. So, the DGFiP intends to extend this same principle to garden sheds, verandas, sheds and other garages nestled at the bottom of a private property. AI, a tax controller with a future?
The reality is more nuanced. The detection of fraud thanks to this technology is in truth “only in its beginnings”, tempers the senator of Jura Sylvie Vermeillet (Centrist Union), author of a detailed report on the subject, published in early April. After two years, the device on the swimming pools is still not running at full speed. Corsica and Overseas – highly “piscinable” departments – have not yet been scanned by the eyes of AI. The development, it should be noted, cost 24 million euros. Which, compared to the 40 million recovered by the State thanks to the system to date, makes the return on investment rather modest for the moment. The rise in power on the other buildings is just as difficult to believe at the moment.
“Far from state-of-the-art technology”
Because the task promises to be more difficult: the shapes are more heterogeneous, the constructions to be uncovered less colorful. At the start of the swimming pool project, tarpaulins, parking spaces painted blue or simply above-ground pools could trigger image recognition. The public finance unions initially mentioned 30% errors on the part of the machine. A rate that the DGFiP says it has significantly reduced since then. “Only 6% of swimming pool owners contacted after detection by AI dispute.” However, only around ten data scientists run the machines. Asked about new recruitments underway within the service, the DGFiP has not yet responded to this question. For comparison, in the United States, the local tax authorities, the IRS, are currently hiring nearly 20,000 new profiles, several dozen of whom, according to advertisements available online, are dedicated to AI. To date, some $375 million in fraudulent transactions have been uncovered by AI across the Atlantic.
More broadly, in France, the use of artificial intelligence in the fight against fraud among individuals and businesses seems a bit overestimated. Current progress in the field is, in reality, mainly that of data mining – this crossing of multiple tax, customs or even judicial data – within applications like Galaxie, also launched in 2022 within the French tax authorities. “A few years ago, the administration had difficulty cross-checking very simple errors, discrepancies between professional and personal declarations, particularly when they were made in two different departments, which is often the case for independent workers working in Paris and living on the outskirts”, notes tax lawyer Arnaud Tailfer, partner at the Arkwood firm. Today, 50% of controls in France are detections of this data mining. More recently, the information provided by taxpayers themselves on social networks has also been used by the tax authorities to ferret out fraudsters.
The real contribution of artificial intelligence to this effort may not be what we think. “If the data mining division uses classic machine learning, it never uses deep learning, although it is at the origin of most recent progress in AI,” points out the senatorial report written by Sylvie Vermeillet with her colleague from Isère Didier Rambaud (Renaissance). Currently, AI dedicated to fraud only relies “on relatively basic models, far from the state of the art of technology (without even talking about generative AI), and far from what is commonly done in large companies to meet their business needs, ranging from oil prospecting to financial services, including online advertising or risk modeling on a network infrastructure (SNCF, RTE, etc.)”, continue the elected officials .
VAT, IFI…
The latter suggest testing other methods, such as “unsupervised learning”. This sub-domain of AI would make it possible “to reveal fraudulent, unusual, complex, or even unknown behaviors or arrangements, by itself establishing statistical links and correlations – sometimes unsuspected – between the elements”, details the senatorial report. Especially since the State does not lack data to progress. Innovate. The administration would have no less than 6 terabytes, or 6000 GB of useful data linked to tax fraud to train new programs, she recently boasted to franceinfo.
Beyond swimming pools, many areas could benefit from it. “VAT, the first source of fraud,” observes Sylvie Vermeillet. Cryptocurrencies, as pointed out by the Minister of Public Accounts Thomas Cazenave in his new anti-fraud plan. Tax lawyer Arnaud Tailfer believes that AI can help fight against evasion of the IFI, the real estate wealth tax. “This lends itself well to so-called predictive models,” he explains. Concretely, by analyzing recurring fraud patterns using AI, then cross-referencing them with other data – the price per square meter per city or neighborhood giving a good map of the type of owners, the amount of property tax as well as than the level of income – it would be simpler to concentrate the tax authorities’ efforts on a few profiles. And perhaps make its controls both more precise and more efficient. The Court of Auditorsin a note dated last fall, says nothing else, calling for this type of control to be strengthened using artificial intelligence, via “the exploitation of an increasingly large number of data”.
When? Hard to say. Three obstacles persist. A reliable estimate of tax fraud does not exist in France. Therefore, the potential effectiveness of AI remains debated: will it have a profound effect, or will it scoop up the ocean with a teaspoon? There are also internal reluctances, within the State. “If the DGFiP is proactive on AI, I have not felt this ambition in the social sphere, within Urssaf, Caf or pension funds,” explains Sylvie Vermeillet. Finally, an AI that is effective very quickly could prove problematic somewhere. As counterintuitive as it may seem, this is what the senators are talking about, through the example of the fight against drug trafficking through customs. Today, AI can already be very useful in detecting a bale of cocaine in a container. But “if tomorrow a technology made it possible to detect a hundred times, a thousand times more traffic, the administration would simply not have the means to cope with the scale of the task, at least with constant law and unchanged organization” , estimate the elected officials. AI would thus have economic consequences contrary to those hoped for. “This gain in efficiency would be paid for directly by a massive loss of attractiveness for French ports, in favor of destinations where controls are more permissive.”
.