This post is the first contribution to 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.

In a 1990 episode of the sitcom Roseanne, Roseanne and Dan Conner (played by Roseanne Barr and John Goodman) are frantically working on their joint Form 1040 on the April 15 due date, when they are stymied by uncertainty about whether Roseanne received and lost a Form 1099 reporting the $400 she earned last year selling magazine subscriptions, or whether a 1099 would not have been issued for so small an amount.¹  Thinking that they will include the $400 on their return only if the IRS had been informed of it by a 1099 filed by the payer, they attempt to determine the minimum payment to an independent contractor for which a 1099 is required. Finding nothing on point in the 1040 Instructions, and getting a busy signal when calling the IRS helpline, they make a trip to their local IRS office.  After their long wait in line, an obnoxious IRS employee (“you pencil-pushing geek,” according to Roseanne) tells them a 1099 is required only for payments of $600 or more. Back home, they finish their return without including the $400. 

If a twenty-first-century version of the Conners had the same question, they might skip the printed instructions, the unsuccessful phone call, and the trip to the crowded IRS office, and go straight to the IRS’s “Interactive Tax Assistant” (ITA).² On ITA’s homepage they would read: “You can use this tool to get answers to your tax questions. Choose a topic, then enter basic information to find your answer.”  The IRS’s ITA is a prominent example of interactive automated legal guidance provided to the general public by a federal agency. Other examples include “Emma, Our Virtual Assistant” of the United States Citizenship and Immigration Service (USCIS),³ and the Federal Student Aid Office’s “Aidan,” “a virtual assistant that can answer your questions about federal student aid.”⁴

Joshua D. Blank and Leigh Osofsky are the leading scholars of the emerging field of agency-provided automated legal guidance. They have now published the foundational book on the subject, Automated Agencies: The Transformation of Government Guidance.

I. Interactive Guidance: What It Is (and Isn’t)

Focusing on ITA, Emma, and Aidan, Automated Agencies offers the reader an under-the-hood look at automated agency guidance. The authors explain that, after a taxpayer consulting ITA has selected from a menu of general topics, the taxpayer is presented with a series of questions, which must be answered by selecting from a limited menu of possible responses.⁵ This typically continues for several iterations, with each new question and menu of responses depending on the answers to the previous question.  Eventually–typically after only a few minutes–ITA gives the taxpayer a clear, user-specific, bottom-line answer, along the lines of “based on the information you have provided, your expenses for X [are/are not] deductible.” Under the hood ITA is decidedly low-tech. In substance it is nothing more than an automated decision tree, and decision trees in printed form go back to 1963.⁶ Although the average taxpayer would probably find consulting ITA easier than tracing one’s route through a printed decision tree, the more user-friendly interface is the only difference between ITA and a printed decision tree.

Emma and Aidan are fundamentally different from ITA, but similar to each other. Aidan allows the user to ask any question whatsoever, in the user’s own words; Aidan uses artificial intelligence and natural language processing to interpret the questions.⁷ While Aidan is much more open-ended than ITA in the user inputs to which it will respond, the downside is that it does not give a definitive user-specific answer.  For example, as Blank and Osofsky relate, the question “How can I discharge my student loan?” results in a brief general explanation of the circumstances under which loans can be discharged and a link to an article on the subject, but Aidan makes no attempt to determine the user’s particular circumstances or whether those circumstances qualify for discharge. Emma is a variation on the same theme.

By the technology standards of 2025, neither the decision tree approach of ITA nor the natural language approach of Emma and Aidan is cutting-edge technology. Despite the widespread availability of generative AI–such as the free version of OpenAI’s ChatGPT available at https://chatgpt.com–Blank and Osofsky report that no federal agency is currently using, or has indicated it is attempting to develop, generative AI to interpret a user’s questions and to generate text providing a definitive answer customized to the user’s circumstances.⁸

II. The Future of Automated Guidance: Less than Meets the Eye?

The focus of Blank and Osofsky is on what today’s automated agency guidance does.  Given the state of technology in 2025, however, what automated guidance does not do is at least as notable. As noted above, the substance of the IRS’s ITA is simply a decision tree, and decision trees have been around in paper form for over sixty years. As for Emma and Aidan, in essence they are just specialized versions of Google searches. Using natural language processing (like a Google search) they scour their limited databases (relating to immigration law in the case of Emma, and to financial aid law in the case of Aidan) and produce possibly relevant documents for the user to peruse. What is conspicuous by its absence is any attempt to use generative AI (such as ChatGPT) to produce original text answering users’ queries. With free ChatGPT available to anyone with an internet connection, the explanation for the absence is obviously not the unavailability to federal agencies of generative AI technology. As Blank and Osofsky explain, “Agencies have not yet embraced automation technology that produces its own content in response to users’ inquiries,” out of concern that generative AI “could result in users receiving ‘hallucinated’ answers or other deviations from the underlying law.”⁹

The problem is not so much that users who receive incorrect user-favorable responses from an agency’s chatbot could use those responses to avoid penalties for noncompliance, or even to benefit from a chatbot’s deviation from substantive law. As Blank and Osofsky explain, under current law a taxpayer cannot hold the IRS to incorrect taxpayer-favorable informal agency guidance—not for purposes of determining the taxpayer’s tax bill, and not even for the more limited purpose of penalty avoidance.¹⁰

The real problem–for which a legal rule prohibiting user reliance is no solution–is the potential for agency humiliation in the unforgiving court of public opinion. If a taxpayer today poses a tax question not to ITA, but to ChatGPT, receives a hallucinatory and obviously wrong answer, shares it on social media, and the answer goes viral, OpenAI (the proprietor of ChatGPT) would not be greatly harmed by the fallout. Everyone understands that chatbots at their current stage of development are subject to hallucinations and other errors, and that ChatGPT is a generalist rather than a tax specialist. By contrast, public opinion would not cut the IRS any slack if its government-sponsored tax-specialist chatbot made the same mistake.  The chances are practically nil that risk-averse officials of any federal agency would offer the public an agency chatbot, given the current state of generative AI technology. And so we are left with two decidedly less-than-cutting-edge types of interactive agency guidance–in essence, decision trees and narrowly specialized database search engines.

III. Roseanne and Dan in 2025

Recall the 1990 Roseanne episode described earlier. On April 15 Roseanne and Dan Conner, finding no answer to their urgent tax question in the printed Form 1040 Instructions, and despairing of the IRS answering their phone call, as a last resort take their question to their chaotic local IRS office. What might Roseanne and Dan have done if they were transported to April 15, 2025, and were faced with the same question about whether Roseanne should have received a Form 1099? They would have had available, of course, a wealth of online resources undreamt of a quarter century earlier, including both ITA and all IRS forms, instructions, and publications. As it happens, the answer would have been exactly the same as it was in 1990. Despite $600 in 1990 being the equivalent of about $1,500 today,¹¹ the threshold in 2024 for a payer being required to file a Form 1099 remained $600 per payee.¹²

If they were aware of ITA, that would have been the obvious resource to consult first. They would, however, have been disappointed by ITA. A search for “1099″ in the “Interactive Tax Assistant Search” box on the ITA home page would have produced five results–four relating to retirement savings, and one to debt forgiveness–none of which address the rules for 1099s for non-employee compensation.¹³ They might next have consulted, still on the IRS website, the Instructions for Form 1040. But they would again have been disappointed, as the information they sought was not to be found in the 1040 Instructions for 2024.¹⁴

Having failed with both ITA and the Form 1040 Instructions, they would probably have despaired of finding an answer on the IRS website, and would have turned to other online resources.  If they had done a Google search for “should I have received a Form 1099,” the first hit would have been an article from Investopedia, “What are 10 Things You Should Know About 1099s?” Google would have helpfully displayed the key sentence from the article: “Businesses are typically required to issue a 1099 form to a taxpayer (other than a corporation) who has received at least $600 or more in non-employment income during the year.”¹⁵ Or, if they had wanted to use the most technologically advanced free service, they could have posed the question to OpenAI’s free version of ChatGPT.  They might have asked, “Should I have received a 1099 for income from selling magazine subscriptions?” If they had, they would have received the response, “If you sold subscriptions as an independent contractor (like door-to-door, on commission, or through a side gig), and you earned $600 or more from a single payer (such as a company or platform, then yes, you should have received a 1099-NEC.”¹⁶

What lessons might be gleaned from this exercise in updating Roseanne? First, that the subject matter coverage of ITA is very spotty. ITA does not purport to answer many common taxpayer questions; the $600 1099 threshold is just one of hundreds (maybe thousands) of non-esoteric tax questions not addressed by ITA. Second, that Google and OpenAI have rushed in where the IRS fears to tread, with both a search engine based on natural language processing (Google Search) and generative AI (ChatGPT). And, at least in this instance, both non-governmental tools have performed admirably. Of course, neither now nor in any imaginable future will taxpayers be entitled to rely on either Google Search or ChatGPT wrong answers for purposes of either penalty relief or determinations of their tax liabilities. But if the results of their searches and inquiries are, in fact, highly reliable, the absence of legally protected reliance will be a moot point with respect to the overwhelming majority of responses that are accurate.

This is the picture today, and the picture is not likely to change in the near future. Risk-averse agencies unwilling to take responsibility for their own hallucinating chatbots will continue to offer automated interactive guidance only in the relatively primitive forms of automated decision trees and limited-database search engines. Taxpayers will find interactive agency guidance less helpful than readily available free non-governmental alternatives, and automated agency guidance will increasingly find itself in the shadow of Google Search, ChatGPT, and the like.

Automated Agencies does not address what I believe will be the next big question in the evolution of automated guidance–how agencies can and should revise their automated guidance in response to competition (so to speak) from non-governmental forms of automated guidance. I suspect, however, that they will address that question in their future work. If they do, their analysis of this next big question will be on the top of my reading list.

Lawrence Zelenak is the Pamela B. Gann Distinguished Professor of Law at Duke Law School.

---

¹ Roseanne, “April Fool’s Day,” first broadcast April 10, 1990 on ABC. Directed by John Pasquin and written by Roseanne Barr, Stephen Paymer, and Matt Williams.

² irs.gov/help/ita. 

³ uscis.gov/tools/meet-emma-our-virtual-assistant.

⁴ studentaid.gov/aidan.

⁵ Pp. 46-47.

⁶ James M. Morgan & John A. Sonquist, “Problems in the Analysis of Survey Data, and a Proposal,” J. of the Am. Stat. Ass’n 415 (1963).

⁷ Pp. 42-43.

⁸ Pp. 103-04, 207.

⁹ P. 207. See also pp. 103-04 (making the same point).

¹⁰ Pp. 175-77.

¹¹ Bureau of Labor Statistics, CPI Inflation Calculator, bls.gov/data/inflation_calculator.htm.

¹² IRC §6041(a) (prior to amendment in 2025). In 2026 the threshold will increase to $2,000. IRC §6041(a).

¹³ irs.gov/ita-index-search?search=1099, visited April 15, 2025.

¹⁴ Internal Revenue Service, 1040 (and 1040-SR Instructions), Tax Year 2024.

¹⁵ Google search conducted April 15, 2025; Robert W. Wood, “What are 10 Things You Should Know About 1099s?,” available at investopedia.com.

¹⁶ Result of ChatGPT search conducted on April 15, 2025, at chatgpt.com (bolding in original).