Court documents released in August revealed that Swiss tax officials are investigating art dealer and freeport magnate Yves Bouvier for allegedly concealing CHF 330 million in profits. The Swiss authorities believe that Bouvier used a fictitious residence in Singapore to evade taxes in his home country, and confiscated one of Bouvier's properties, reportedly worth CHF 4.5 million, as a pledge while they continue investigating his finances.
The investigation, however, was nearly derailed in its early stages due to a single vulnerable tax official. An escort girl known only as Sarah has testified that in September 2017, Yves Bouvier sent her to a conference to seduce a key official with Switzerland's Federal Tax Administration. Sarah's honeypot adventure took place mere months after Swiss tax officials had begun looking into Bouvier's finances.
Yves Bouvier's plan didn't work out-Sarah was caught stealing mail and spent sixteen months in prison-but it uniquely highlights how vulnerable tax oversight remains to human error and its potential exploitation by malicious interests. And it adds to the growing pile of evidence that artificial intelligence and machine learning represent "the new frontier" in tax administration.
Technological solutions to ferret out tax cheats
The idea of using artificial intelligence as an impartial observer to spot those hoping to shield their income from the long arm of the law isn't new-as early as 2015, researchers at MIT had crafted an algorithm which could flag a certain kind of tax shelter. The concept has caught on with a number of governments in the past year, and more governments are sure to follow suit as they try and plug the holes that the pandemic has left in their national coffers.
Denmark, which lost nearly €2 billion to tax evasion in 2018, implemented artificial intelligence tools which successfully identified 85 out of every 100 cases of tax evasion. France, meanwhile, passed a law as part of the country's 2020 budget allowing tax authorities to deploy algorithms to trawl through social media and detect signs of tax evasion, smuggling and undeclared income.
India is hoping to recoup more than €34 million a year by rolling out a machine-learning tool, developed by two Indian researchers at the University of California Berkeley, which can identify bogus firms set up purely for tax evasion purposes. The technological solution is particularly attractive to New Delhi given that it has had problems in the past with human tax investigators-last year alone, India had to sack 85 tax officers, including 64 high-ranking officials, after reports of widespread corruption and bribe-taking.
Japan, meanwhile, is hoping to bring an AI tax inspection system online as early as next year to reverse a worrying uptick in companies failing to declare all of their taxable income. In addition to analysing financial reports, Tokyo's AI system will comb through voice recordings of company executives describing their firm's financial performance to pick up on behaviours previously associated with illicit financial reporting.
Earlier this summer, meanwhile, a team of researchers from Mexico and Hungary used artificial intelligence and network science to estimate that Mexico loses an average of 60 billion pesos (€2.3 billion) a year due to VAT evasion using fraudulent digital tax receipts. The researchers' AI-based modelling also found that the amount evaded is steadily increasing-in 2015, only 40 billion pesos (€1.5 billion) slipped through the cracks, compared to 77 billion (nearly €3 billion) three years later. The algorithm deployed by the researchers not only shed light on the scale of the problem, but was able to identify patterns in tax evaders' behaviour-raising hopes that the results could be used to train a neural network to detect other similar cases.
Building a more efficient tax scheme
AI-based solutions offer governments a way to crack down on tax evasion while using limited resources and sidestepping problems of human error and corruption. But the technology can be used to streamline other parts of a country's tax system as well. For example, a team of researchers from Harvard and the cloud computing firm Salesforce have developed an "AI Economist" which can propose new tax policies.
The technology underpinning the AI Economist is quite innovative. Most previous attempts to model tax policies have relied purely on past experience to predict how people will respond to policy changes. The AI programme developed by Harvard and Salesforce, however, uses reinforcement learning-a subset of artificial intelligence that has helped AI model increasingly complex behaviours-in order to craft the most optimal balance between equality and productivity. The AI Economist uses reinforcement learning in two ways-simulating how people would react to different tax options, including how they would seek to dodge tax, and evaluating the global impact of various tax proposals.
The programme could provide policymakers with an invaluable tool by constantly tweaking policy options to maximize both economic efficiency and equity and doing what no human can do-simulating millions of years of economics. The AI tool-which could even take into account pandemic-related restrictions such as social distancing and remote working-could be particularly useful as governments try and patch the holes which lengthy lockdowns have left in their budgets. A number of countries, including the UK, are contemplating substantial tax rises to combat the economic downturn; a tool like the AI Economist could help ensure that tax policy changes pad government coffers without jeopardising the country's economic recovery.
Many AI-based tax tools, including the AI Economist, are still very much in their infancy. Nevertheless, they have the potential to revolutionise the sector by sharply curbing the possibility for human error or undue influence and allowing for large-scale modelling.
