This post is part of Notice & Comment’s symposium on Joshua D. Blank and Leigh Osofsky’s Automated Agencies: The Transformation of Government Guidance. For other posts in the series, click here.

It is a pleasure to participate in this symposium on Josh Blank’s and Leigh Osofsky’s excellent new book, Automated Agencies. Their work intersects in significant and interesting ways with my own work on tax administration and democratic theory, and I will focus on their contributions at that nexus. But first, I want to comment on the remarkable prescience of this project.

When Professors Blank and Osofsky began their work on “simplexity” and later began to connect it to “chatbots” I viewed the problem—myopically and through luddite-shaded glasses—as narrow matter. As they defined it, simplexity arises when “the government presents clear and simple explanations of law that is, in fact, ambiguous and complex” (AA at 5). To be sure, this is a longstanding and significant challenge in administering the federal income tax. The Tax Code is notoriously complex, and in contrast to other thorny areas of the law, essentially every single American citizen or resident is subject to the income tax laws and is personally responsible for complying with those laws every year.

To me, back then, the roots of the problem seemed more interesting and, frankly, a more tractable target for remedying simplexity than did confronting and attempting to improve the IRS guidance that exhibited simplexity. As Blank and Osofsky explain, various “customer service mandates” already required the IRS and other agencies to explain legal obligations in plain language, and extreme complexity was already a primary concern of those mandates and the measures the IRS undertook to comply with them (AA at 58-62).

I saw two plausible—if admittedly challenging—solutions to simplexity.

First, 40 years after the Tax Reform Act of 1986, the Tax Code had again grown too complex. The time seemed ripe for a significant reform that broadened the base and eliminated opportunities for legal and illegal tax avoidance; the problem of simplexity could be reduced by confronting the underlying complexity of the tax system. While I expected to disagree with the first Trump Administration’s distributional preferences, I took some solace in the hope that tax reform seemed to be on the agenda.

Second, after more than a decade of pinched budgets, it was clear by the mid-2010s that the IRS needed more funding and more personnel to fulfill its customer service obligations.

I would have said that both solutions—simplification and real investments in administration and enforcement—shared broad support among tax law scholars, policymakers, and the general public. Together, reform and a reinvigorated tax administration would go a long way toward mitigating and circumscribing the conditions that give rise to simplexity.

A decade later, the world looks very different, and the project that commenced with Blank and Osofsky’s insight on simplexity looks more prescient than ever. Substantive simplification, it turns out, was a pipe dream – the partisan whipsaw pattern that developed in the Bush and Obama Administrations became even more severe under Trump and Biden (generally this whipsaw consists of tax cuts at the top and special business provisions in the Republican administration, and special tax credits and tax cuts targeted at the least well-off in the Democratic administration). In recent years, elected leaders from both parties have added enormous complexity to the Tax Code, including a variety of rules that impose complexity on regular taxpayers doing regular things (that is, not just on sophisticated taxpayers who have the resources and wherewithal to hire tax professionals to help them do things that are complicated for non-tax reasons).

The examples are many, but a few of the most far-reaching and concerning include small business owners using pass-through entities (§ 199A), consumers upgrading their houses with energy efficient improvements (§ 25C, § 25D), and families claiming the Child Tax Credit (§ 24). Each change or addition to the Tax Code might be justifiable on its own narrow terms, but together the constant changes have contributed to increasing confusion and frustration, which I see constantly in my work helping low-income people prepare their tax returns and representing them in disputes with the IRS. It’s almost impossible to tell self-employed small business owners prospectively whether they will qualify for the § 199A deduction or why they might or might not; the technical details of energy efficiency tax credits caused many taxpayers to expect tax benefits that didn’t pan out (sometimes because they didn’t meet the technical requirements, sometimes because the credits were non-refundable); and the Child Tax Credit amounts and refundable amounts changed almost every year since 2019, leading to confusion and frustration.

Over the same period, the possibility of a reinvigorated, customer-service-oriented IRS waxed and then waned completely. The basic investments in technology and hiring that the Biden Administration initiated might have been the fruits of a bipartisan compromise in an earlier political era. But it was enacted with only Democratic support, and was turned into a political bogeyman under withering attacks from Republicans, culminating in the second Trump Administration systematically gutting the entire effort.

Today, the federal income tax is more complex than ever before, with new, incredibly complicated rules that affect the tax liability and reporting obligations of tens of millions of regular taxpayers adding to the already severe challenges that existed back ten years ago. Further, there no longer seems to be any political constituency for tax reform in the traditional sense (broadening the base and reducing rates), and the IRS is more hobbled than ever, with little hope for meeting the basic demands on the system through more and better-trained personnel.

In short, the conditions that foster simplexity have become even more deeply ingrained. A decade after Blank and Osofsky coined the term, simplexity perfectly captures an increasingly pervasive dilemma: the IRS frequently produces guidance that elides the intricate details of various tax rules, because that is the only way to make those rules remotely understandable for the lay audience that is subject to those rules. The work that Blank and Osofsky have undertaken over this period essentially anticipated this state of affairs, and importantly it also anticipated the major technological development over this same period: the dawn of generative, conversational artificial intelligence (AI).

In Automated Agencies, Blank and Osofsky bring all of this work together with an eye toward navigating through the AI era. As the authors note, thus far the application of AI in federal government automated legal guidance—in particular in tools like the IRS’s Interactive Tax Assistant (ITA), as well as USCIS’s Emma and the Department of Education’s Aidan—has been limited to using AI to interpret inputs and filter through predetermined outputs, rather than to “generate spontaneous text to answer questions” (AA at 48). But surely generative AI responses are just around the corner, as the technology permeates the private sector and the public comes to expect it. This brings true urgency to Blank and Osofsky’s project.

The stakes are high, and while the tax system is just one of the examples that Blank and Osofsky survey, it is a particularly dire one from a democratic perspective. An unexpected effect of automated legal guidance is that “[w]hile, at first blush, they appear to make legal compliance obligations easier, they also exacerbate a two-tiered legal system” (AA at 121). While well-heeled taxpayers can pay lawyers and accountants to plan their affairs and ensure compliance with the law—or at least penalty protection if they run afoul of tax authorities—regular people are increasingly relying on automated legal guidance, and so are more likely to fall into the simplexity traps of guidance diverging from underlying law and relying on guidance that the agency may not—or cannot—actually stand by.

Thus arises the most fundamental democratic challenge in automated legal guidance: what the authors adeptly identify as “unequal access to justice” (AA at 144). Blank and Osofsky provide many intriguing examples, including a particularly poignant one in which a hypothetical “chronically ill taxpayer” might query IRS’s ITA about whether she qualifies for a deduction for payments for an in-home health aid (AA at 148). Despite statutory language and legislative history that, in the hands of an experienced tax professional, clearly establish that such care can be deductible, ITA provides a more general answer that household help is not a deductible medical expense (AA at 149). This may be correct for “the vast majority of taxpayers,” but not for their hypothetical individual (AA at 151).

What to do about sometimes-incorrect automated agency guidance? Of course, there already exists a well-developed set of legal constraints on certain federal agency communications about the law, in the form of the Administrative Procedure Act. Blank and Osofsky wrestle with how to fit ITA responses into this framework—an ITA statement might be treated as an interpretive rule that is not subject to the procedural requirements of a legislative rule such as notice-and-comment rulemaking. But “there is an inherent, and problematic contradiction in characterizing the statement as interpretive, or saying that interpretive, but incorrect, statements are not subject to procedure” (AA at 150). If ITA guidance represents anything less than a “final agency action,” there is no judicial recourse when ITA statements mislead. The authors make a convincing case that ITA responses and other similar types of automated agency guidance “may simply have no place at all within the administrative law framework” (AA at 152).

The current trajectory of automated legal guidance, and the poor fit of the APA framework as applied to such guidance, suggests a future of unconstrained government “slop.” The IRS will almost certainly increase its use of automated responses or generative AI to provide taxpayers with individualized information, but without democratically minded controls, there is little hope that the result will be improved government services.

Blank and Osofsky call for a return to first principles of administrative law, arguing for the need to make automated agency guidance attentive to transparency and accountability as a route toward democratic legitimacy. These are the core democratic values that administrative law scholars have long emphasized, but even in entirely human-animated bureaucratic systems they have proved challenging to realize in practice. As I have documented, the tax system has been plagued by a lack of public participation in rulemaking, even as APA procedures have been treated as a source of democratic legitimacy. Judicial review of ITA responses—even in their current form, let alone a future iteration of generative ITA—is hard to contemplate. (Blank and Osofsky agree that simply applying these standard solutions—more formal procedures and/or expanded judicial review—would be unsatisfactory (AA at 191).)

This brings us to Blank and Osofsky’s excellent and helpful new contribution in this book, building on their 2022 ACUS study: a set of precise prescriptions that would, together, protect against the many pitfalls that exist now and will only grow more problematic in the near future. They call for reforms to make agencies take automated guidance more seriously by making it binding on the agency and, in the IRS context, allowing reliance on certain agency guidance to afford penalty protection in a similar manner to when a sophisticated taxpayer relies on a well-founded opinion letter from a lawyer (AA at 175-76).

Their strongest points, in my view, are around transparency. They call for automated agency guidance to include disclaimers about accuracy, and disclosure when the underlying law is ambiguous or unsettled; agencies to provide explanations when automated guidance is altered prospectively; and agencies to disclose the “decision-tree” structures that underlie automated responses (AA at 171-74, 178).

Finally, Blank and Osofsky call for greater accessibility to “human customer service representatives” (AA at 184). This got me to thinking: in the IRS, human employees are, in fact, subject to precisely the sort of transparency-minded procedures that Automated Agencies contemplates for ITA: the Internal Revenue Manual is quite literally an instruction manual for administering the Tax Code. The I.R.M. lays out, in sometimes painful detail, how IRS employees are meant to parse and navigate essentially every issue that might arise in tax administration, including how to evaluate substantive legal issues that give rise to simplexity in IRS publications. Tax practitioners often use the I.R.M. as a resource to determine how the IRS will approach a particular issue for a particular taxpayer.

Teaching humans to use the manual—i.e., how to administer the tax system we have, as Kristin Hickman memorably described the challenge—is a significant undertaking. A 2023 report by the Treasury Inspector General for Tax Administration detailed the vast resources required for basic education of IRS revenue agents. Training a single agent in the Small Business/Self-Employed division takes two to three years. The report estimated that training for every 100 new agents in that division requires a full-time commitment from 18 experienced revenue agents over a two-year period. These “on-the-job” instructors are taken fully offline in order to train new hires—that is, the training commitment is so intense that they are not expected to conduct regular casework while assigned as instructors. More specialized training, for issues related to large businesses and high-income individuals, requires even more time and effort.

To my mind, this is suggestive of the best “fix” for the simplexity conundrum in light of the potential power of emerging AI technologies: harnessing generative AI by putting it in the hands of human interlocutors. Perhaps there can be some future return to sanity and humanity, in which a more adequately funded tax administration apparatus can use a well-trained future iteration of ITA to facilitate human-provided customer service. Wishful thinking, perhaps. But Blank and Osofsky’s work suggests to me that the best future for automated agencies must include a renewed commitment to the human touch. (Beyond training, any tax professional who has contacted the Practitioner Priority Service hotline on behalf of a client can appreciate that reducing the “5-7 minute” intervals of waiting on hold while an agent consults the I.R.M. can add up quickly.)

Automated Agencies is a timely, important, and truly thought-provoking book that reaches far beyond even the vast reach of the tax system. In hindsight, my own focus on substantive tax rules and tax administration procedures was never going to solve the simplexity problem. Even if we weren’t standing on the precipice of bringing the generative AI revolution directly to bear on government communications, the reality of legal complexity is inescapable in this complex society. Blank and Osofsky have been at least one step ahead, and I’m hopeful that this book can mark the beginning of a coherent approach to automated agencies.

Clint Wallace is a Professor of Law at the University of South Carolina Joseph F. Rice School of Law.