Since 2014, Ben Alarie and his team at Blue J Legal have worked to apply machine learning (ML) principles to the process of tax advising (among other areas of law). Through a series of articles in Tax Notes Federal, Alarie and his coauthors provide a window into their artificial intelligence prediction engine. Their commentary is crucial: big data has arrived in legal and accounting practice, and some degree of transparency may improve tax equity and administration. In addition, these articles yield important and interesting insights about various doctrines in tax law.
In winter 2022, Alarie and his coauthors gave us three short articles: a general review of ML’s potential in tax practice and two applications of Alarie’s ML model to existing controversies.
These case studies involve the temporal scope of the step transaction doctrine (as implicated in GSS Holdings (Liberty) Inc. v. United States) and whether activities rise to the level of a trade or business (at issue in Olsen v. Commissioner). In both cases, the trial courts ruled against the taxpayers, and both taxpayers currently are appealing these decisions. Alarie and his coauthors evaluate these taxpayers’ positions on appeal—including these taxpayers’ likelihood of success.
These articles (and Alarie’s Tax Notes column more generally) emphasize the specific utility of today’s ML in the production of legal work. Alarie and his coauthors allude to two distinct contexts in which tax advisors may employ ML: during ex post compliance and controversy, and during ex ante planning. In both contexts, ML excels with issues that involve “a sizable body of case law and a fact-intensive inquiry” (1236). Drawing on the extensive data generated by these issues, Alarie’s model weighs multiple elements and computes cross-correlations rapidly and with quantitative precision. This deep and dynamic functionality may “uncover hidden statistical patterns” that shape litigation strategy or reveal infirmities in proposed transactions (662). For ex ante analysis, the rewards are efficiency and effectiveness, in that advisors can focus on the facts and factors most likely to be dispositive for their clients. For ex post analysis, the benefits primarily arise from increased certainty, either through the quantification of tax risk or restructuring that avoids significant pitfalls.
Alarie and his coauthors emphasize the integral role of human advisors in workflows that incorporate ML. In constructing and deploying Alarie’s model, expert humans play a significant part. These natural persons identify the legal questions that the model addresses, then translate the relevant primary authorities into “structured data” that the model can use (663). The model’s predictions require human interpretation, both to apply the model’s quantified legal framework to the instant facts and to generate appropriate argumentation in light of the model’s results. There are humans in the loop, and their “skill and judgment” matters (1238). For Alarie and his coauthors, ML is not (yet) a threat to the legal or accounting professions. Indeed, they see “synergy between technology and the tax professional” that enhances those advisors’ productivity and—perhaps—also the quality of their professional lives (1238).
One might ask, of course, whether Alarie’s model (neutrally) provides “better information” to advisors, or whether the model’s existence changes the process of legal development more fundamentally (664). Within the field of taxation, longstanding norms have facilitated the pooling and dissemination of knowledge not unlike that generated by Alarie’s model. The advent of ML implies that associates’ archived analog case charts can be recreated in summary form with the push of a button, and a chatty phone call to a seasoned colleague becomes a keyboarded query into Deep Blue. Although ML may proletarianize taxation by deemphasizing historical networks and relationships, the idiosyncratic aspects of legal practice—moments of creativity, deep insight from engagement with primary authorities—also risk marginalization. Similarly, ML inherently incorporates biases in training data and algorithm construction (and, to be fair, Alarie and his group appear very aware of the potential for these biases). Communities of humans carry their own biases, of course, but ML may remix or augment these biases with unpredictable effects. Overall, Alarie and his coauthors emphasize balance between the human and machine aspects of advising. More should be said, however, about ML’s implications for the substance of law and practice going forward.
For example, ML risks converting fuzzy standards into something more like bright-line rules. As Alarie and his coauthors note, ML shines in precisely these circumstances, guiding planners ex ante and emphasizing critical facts in ex post controversies. The effects—saved time, reduced uncertainty—may be salutary. But pernicious results also may follow. Alarie and Di Giandomenico illustrate the power of ML through a fascinating tabulation of how different variables each might affect step transaction analysis in GSS Holdings (1855). The column with the model’s predictions has a flavor of The Price Is Right: how close to 50% can the taxpayer get without dropping below? Standards—and their zone of uncertainty—deter the risk-averse from aggressive tax planning. To some extent, ML converts those standards into discrete variables and numeric outputs that may encourage positions just barely on the favorable side of the quasi-quantifiable line. This shift would pressure enforcement, among other things. Furthermore, in the controversy context, ML may entrench particular legal understandings at the expense of open-textured inquiry. Judges rely on parties’ advocates (as well as their clerks) to develop the relevant issues, and a Moneyball approach to briefing ultimately may prove limiting. Alarie and his coauthors consider ML primarily from practitioners’ perspectives, and systemic (or government-side) considerations also should play into any normative conclusions.
Over the last two years, Alarie and his coauthors have provided a wealth of technological and doctrinal insight through their regular columns in Tax Notes. These articles are a significant contribution to the tax literature, as well as larger conversations about artificial intelligence and the law. Policymakers, scholars, and practitioners should attend to this work, and I look forward to more from Alarie’s group in the future.