Machine learning & AI

How AI can uncover corporate tax evasion

Are the terms used in annual reports a key to figuring out how corporations avoid paying taxes?

Words can be just as useful as numbers in spotting corporate tax evaders. Texas McCombs’ most recent study reveals that careful text reading can provide new insights into how businesses attempt to evade taxes, which may not be apparent from financial numbers alone.

Lillian Mills, dean and professor of accounting, and Kelvin Law, co-author from Nanyang Technological University, looked at annual reports from 18 years of U.S. multinational corporations that talked about their business activities in other countries, including tax havens. A total of 183,061 reports were examined by the researchers.

“Using AI to evaluate text data might be a powerful tool for both regulators and investors to uncover business tax evasion,”

Accounting Professor Lillian Mills

Natural language processing, or NLP, was used by the team to look at the text and find patterns and word choices that might show what businesses were doing in tax havens. The PC examination revealed signs about these exercises.

Take, for instance, the scenario in which a pharmaceutical company in the United States has developed a drug that works well to treat heart disease and has a high profit margin. The drug’s specific formula is protected by intellectual property (IP), and the company claims to have “established a subsidiary in Panama to handle manufacturing and production” using the patented formula. The company is able to pay lower taxes through the subsidiary in the tax haven because it routes profits from the sale of the heart disease drug through the use of IP in a nation with low tax rates.

The computer looks for about 80 words, including “manufacturing,” to suggest activities that might be avoiding taxes. “Distributors,” “warehouses,” and “purchasing” are some of the others.

According to Mills, despite the fact that there is no certain way to detect all instances of tax evasion, paying close attention to word choices in an annual report can provide insights that numbers might not:

Added metrics The study uses a new set of measures to determine not only whether a company has a subsidiary in a country with tax havens but also whether that subsidiary is an active subsidiary. When it comes to predicting whether a company is avoiding taxes, the new measures are three times as effective as the previous ones.

operations not disclosed Companies that do not disclose that they operate out of tax havens can be identified using machine learning techniques.

Increased tax evasion. The effective tax rates of nondisclosers identified by machine learning are lower than those of other businesses.

According to Mills, “a powerful tool for both regulators and investors to detect corporate tax avoidance could be using AI to analyze text data.”

“Since they do not have access to the tax returns of businesses, regulators other than the IRS may particularly benefit from that information.” It might direct them to look at publicly available data to find businesses in tax havens that may be using abusive profit-shifting strategies.

More information: Kelvin K. F. Law et al, Taxes and Haven Activities: Evidence from Linguistic Cues, The Accounting Review (2021). DOI: 10.2308/TAR-2020-0163

Topic : Article