Sarah Lawsky (Northwestern) presents Form as Formalization at Duke today as part of its Tax Policy Workshop Series hosted by Lawrence Zelenak:
There are, roughly speaking, two approaches to applying computing to law, which can be thought of as a bottom up approach and a top down approach. The bottom up approach uses large amounts of data in some way—to make predictions, for example. Machine learning is one such approach; neural nets are another. This is generally seen to be the approach with the most potential, the one that leads to the “data driven future” of legal practice. The “top down” approach would derive conclusions from the law itself, after making the law legible to the computer in some way, perhaps through hand-encoding, perhaps through natural language processing or some similar approach to read the actual text of a statute or regulation.
The top-down approach, also sometimes called computational law, is generally considered to have much less potential. But even those whose dismiss computational law point to one example of success in encoding the law: TurboTax and similar tax compliance programs. Thus one finds enthusiastic references to creating “TurboTax for police complaints,”“TurboTax for copyright,” “TurboTax for immigration,” and so forth.
Daniel Katz argues
Tax software companies are essentially selling the simplification of the Tax Code for the user, yet they must provide a product that accurately calculates the user’s tax liability under any scenario and thus must somehow pack all of the Tax Code’s substance into the software program. ... If the software is too complex, for example, it may be very difficult to update the program as Congress changes the Tax Code, as a change in one provision cascades in effect to other provisions. It would be in a tax software company’s interests, therefore, to develop a program that is no more complex than needed to produce accurate user tax liability calculations. Perhaps a good measure of Tax Code complexity would then be the complexity of reliable tax compliance software. [J.B. Ruhl & Daniel Martin Katz, Measuring, Monitoring, and Measuring Legal Complexity, 101 Iowa L. REV 191, 196 (2015).]
Such characterizations of TurboTax and other similar programs are incorrect. Tax preparation programs do not encode the law. Rather, as this paper argues, these programs encode the tax forms, which are not law, and which are prepared by the government itself. The difficult part of the coding, and the judgment calls, are almost entirely performed by the government, not by those who code tax preparation software. Tax forms are extraordinary in that they are designed so that people who do not know or understand the law can still comply with the law. Indeed, people may have absolutely no idea why they are filling out certain lines, or what the legal implications of those lines are. And yet they do fill out the forms, and they do thus comply with the law. And tax preparation programs do not simplify the law, because tax forms do not simplify the law. Rather, tax forms collect information and turn portions of the law into an algorithm for taxpayers to apply.
This paper investigates tax forms to learn more about how coding law can be beneficial, and considers what can be learned from coding that could improve tax forms. Part II investigates the relationship among forms, tax preparation software, and the law, and introduces the elements of a form: inputs, outputs, and algorithms. Part III uses the casualty loss rules for individuals, and the recent change in those rules, as an example of the implementation of law through tax forms. As this Part shows, tax forms themselves—not the instructions, but the actual forms, the actual way that entries are set to interact with each other—contain judgments about the law, sometimes law that is unclear. Part IV suggests some possible next steps, including creating more-flexible tax forms and formalizing the law itself. Part V concludes.