Many tax law questions can be thought of as binary classification problems. Algorithms have consistently been shown to outperform human beings in making predictions about these kinds of problems. Drawing on recent advances in machine learning, my colleagues at the University of Toronto and I have created a system that analyzes patterns in the case data for a range of tax law questions to make predictions about how courts will rule in new situations. We report the confidence of our predictions as a percentage based on the probabilistic likelihood of the outcome.
One advantage of our research approach is that it allows us to observe how the probability of a given outcome changes when we alter the fact pattern. In the previous article in this series, we saw how the presence of financial risk factors in a debt vs. equity case can alter the machine learning’s confidence in the predicted outcome. But how well do machine learning algorithms work with more tangible legal questions, such as whether a worker should be classified for tax purposes as an employee or an independent contractor?
Employee or Independent Contractor?
Workers categorized as employees are subject to a number of payroll withholdings, including income tax, social security and Medicare (FICA), and unemployment (FUTA). Employers are responsible for holding these taxes back from their employees’ wages and for paying a corresponding share up to a certain limit. Independent contractors, meanwhile, are responsible for their own taxes. In addition to income tax, they also have to pay a percentage of their self-employment income towards old-age, survivors, and disability insurance, as well as a smaller percentage towards hospital insurance (tax code Section 1401).
With the growth of the gig economy, the employee-independent contractor distinction has received renewed attention in both the media and the courts. While rulings on the status of gig economy workers like Razak v. Uber Techs. Inc. and Lawson v. Grubhub Inc. have focused on the issue of employment status in particular states rather than on federal tax, these decisions necessarily have tax implications for the workers and hirers involved. The consequences of an incorrect classification can be costly, leaving the taxpayer liable for deficiencies and additional penalties—not to mention the cost of fighting the Internal Revenue Service in court.
The IRS’s Factor Tests
In determining a worker’s status for tax purposes, the courts have historically relied on a common-law test codified by the IRS in 1987 as the 20-Factor Test (Joint Committee on Taxation, Present Law and Background Relating to Worker Classification for Federal Tax Purposes [JCX-26-07], 3-5). The IRS factors cover a broad range of issues including training, level of integration in the hirer’s business, control over scheduling, and the right to termination. Taken together, they provide a comprehensive view of a worker’s situation that can be used to determine whether or not an employment relationship exists.
1. Behavioral: who decides when, where, and how the work is to be done?
2. Financial: who is responsible for the costs associated with doing the work?
3. Type of Relationship: what kind of relationship do the parties have?
Of these three categories of factors, behavioral control (the when, where, and how) is often the most important in determining whether a worker is an employee or an independent contractor. But how exactly do the courts weigh the various factors of behavioral control, and how have they influenced the outcome of actual cases?
Recent advances in AI make it possible to answer both of these questions with new levels of precision. As we saw in the previous articles in this series, machine learning algorithms can be trained on data extracted from existing cases in order to make predictions about how a court might rule on a new scenario.
We’ve seen how accurate these algorithms can be when applied to a financial question like debt vs. equity, but how well does machine learning work when applied to the more tangible question of whether an employment relationship exists for tax purposes?
Example: Ramirez v. Commissioner
Let’s look at a recent case that illustrates the key behavioral factors that the courts consider when attempting to determine a worker’s status. Then we’ll explore how the different weights assigned to these factors by a machine learning algorithm can influence the probable outcome.
In Ramirez v. Commissioner, the taxpayer, Juan A. Ramirez, was employed by the Univision network as an on-air personality and radio station manager at KXTN in San Antonio, Texas. When the station encountered financial difficulties in 2005, Ramirez took the initiative to find new sponsors himself. Rather than coordinating these relationships through the station, Ramirez favored a more direct approach. Without input from Univision, he and the sponsors agreed on a fee that would be paid for his services in developing advertising campaigns and promoting the sponsors’ products both on-air and at public appearances.
Although the sponsors were invoiced by Univision for Ramirez’s work, and the network withheld the necessary income and payroll taxes for his other duties at the station, when Ramirez filed his tax return he claimed his $82,000 in income from the sponsorship deals as freelance earnings. The IRS challenged the claim, arguing that Ramirez was not an independent contractor with respect to these deals, but rather an employee of Univision.
At first glance this may seem like a close case, since different factors point to different outcomes. Even though Ramirez made the sponsorship deals himself, his ability to provide his clients with advertising and promotional services depended almost entirely on his employment with Univision as a radio host. In other words, the answer in this case to the question “Who is responsible for the costs associated with doing the work?” is Univision, since they own the radio station. But the answer to the question “Who decides when, where, and how the work is to be done?” is Ramirez: he made the sponsorship deals on his own initiative, and, as the Tax Court emphasized, Univision was really just the conduit through which he received his fees.
The advantages of an AI-powered research approach are especially apparent in a case like Ramirez. While human lawyers and accountants rely on their experience and instinct to judge how a particular combination of factors will sway the outcome in court, machine learning can precisely weigh each factor and evaluate the interactions between them.
The result in this case is a prediction with a confidence level of 95 percent that Ramirez will be found to be an independent contractor, which is precisely what happened. The Tax Court said that, despite Univision’s ownership of the facilities that Ramirez used to broadcast his sponsors’ commercials, “The record does not demonstrate that Univision possessed the requisite degree of control to establish that Mr. Ramirez was acting in his capacity as an employee with respect to promoting his sponsors’ products/services”.
Testing an Alternative Scenario
In addition to control over the manner of the work, our trained algorithm also considers more specific behavioral factors when making its predictions. These include, among other things, whether the hirer or the worker had control over scheduling, whether the work was performed in the hirer’s workplace, and whether the worker was required to submit regular reports. To get a sense of how these other behavioral factors affect the outcome, let’s alter some of the facts of the Ramirez case and see how the changes affect the AI’s confidence in the result.
In our alternative scenario, Univision decides when and where Ramirez will meet with potential sponsors. The network also requires Ramirez to submit weekly reports on his progress in attracting new sponsors. When we run these new facts through the algorithm, we still get the same prediction: independent contractor. The confidence level, however, has dropped to 66 percent. Despite the fact that Ramirez still controls how the work gets done, which the Tax Court called “the crucial test” of whether an employment relationship exists, the outcome has become less certain. And if we imagine that the potential sponsors are picked by Univision rather than by Ramirez himself, the result flips completely—the algorithm now predicts with 68 percent confidence that the court will find Ramirez to be an employee for tax purposes.
Some Conclusions, and Looking Ahead
The key takeaway from this analysis of behavioral control factors in Ramirez is that there is no simple formula for determining whether an employment relationship exists. Enumerating the factors won’t necessarily give us the right answer; even when a crucial factor is present, it may not be enough to decide the outcome. As our alternative scenario for Ramirez demonstrates, all the factors in a case interact with one another, which means that an accurate prediction depends on having a firm grasp of the dynamic relationships between them.
Lawyers and accountants develop this kind of deep understanding through years of experience. The insights of machine learning allow us to augment that professional knowledge in ways that will make the law fairer, more accurate, and more accessible.
The next installment in this series will explore the application of machine learning algorithms to a notoriously challenging tax question: economic substance.
Benjamin Alarie holds the Osler Chair in Business Law at the University of Toronto Faculty of Law and is the CEO of Blue J Legal.